Large language models (LLMs) often produce factually inaccurate content, or hallucinations, which undermines their reliability. Existing factuality evaluation systems usually rely on a single predefined fact source, making them task-specific and hard to extend. We present UFO, a unified framework for factuality evaluation that supports multiple plug-and-play fact sources. UFO integrates human-written evidence, web search results, and LLM knowledge within a single evaluation pipeline, and allows users to flexibly select, reorder, and even define customized sources. The system is accessible through both a Python interface and a web-based demo, offering interactive claim-level verification and visualization. Experiments show that UFO system achieves moderate consistency with human annotations. Overall, UFO serves as a transparent and extensible platform for benchmarking fact sources, comparing LLMs, and enabling real-world fact-checking applications across diverse domains.

As scaling up training data has significantly improved the general multimodal capabilities of Large Vision-Language Models (LVLMs), they still suffer from the hallucination issue, generating text that is inconsistent with the visual input. This phenomenon motivates us to systematically investigate the role of training data in hallucination. We introduce a new benchmark, POPEv2, which consists of counterfactual images collected from the training data of LVLMs with certain objects masked. Through comprehensive evaluation on POPEv2, we find that current LVLMs suffer from training bias: they fail to fully leverage their training data and hallucinate more frequently on images seen during training. Specifically, they perform poorly on counterfactual images, often incorrectly answering “Yes” to questions about masked objects. To understand this issue, we conduct probing experiments on the models’ internal components, revealing that this training bias is primarily located in the language modeling (LM) head, which fails to correctly translate accurate visual representations into textual outputs. Based on these findings, we propose Obliviate, an efficient and lightweight unlearning method designed to mitigate object hallucination via training bias unlearning. Obliviate identifies the discrepancy between ground-truth labels and model outputs on the training data as a proxy for bias and adopts a parameter- and data-efficient fine-tuning strategy that only updates the LM head. Extensive experiments demonstrate the effectiveness of our approach. While only reusing the training data and updating approximately 2% of the parameters, Obliviate significantly reduces hallucination across both discriminative and generative tasks. Furthermore, it demonstrates strong scalability with respect to both model size (2B to 72B) and training data volume, and exhibits promising generalization to hallucination types beyond object-level hallucination.

Recently, large reasoning models have demonstrated strong mathematical and coding abilities, and deep search leverages their reasoning capabilities in challenging information retrieval tasks. Existing deep search works are generally limited to a single knowledge source, either local or the Web. However, enterprises often require private deep search systems that can leverage search tools over both local and the Web corpus. Simply training an agent equipped with multiple search tools using flat reinforcement learning (RL) is a straightforward idea, but it has problems such as low training data efficiency and poor mastery of complex tools. To address the above issue, we propose a hierarchical agentic deep search framework, HierSearch, trained with hierarchical RL. At the low level, a local deep search agent and a Web deep search agent are trained to retrieve evidence from their corresponding domains. At the high level, a planner agent coordinates low-level agents and provides the final answer. Moreover, to prevent direct answer copying and error propagation, we design a knowledge refiner that filters out hallucinations and irrelevant evidence returned by low-level agents. Experiments show that HierSearch achieves better performance compared to flat RL, and outperforms various deep search and multi-source retrieval-augmented generation baselines in six benchmarks across general, finance, and medical domains.

Recently, Chain-of-Thought (CoT) reasoning has significantly enhanced the capabilities of large language models (LLMs), but Vision–Language Models (VLMs) still struggle with multi-step reasoning tasks due to limited multimodal reasoning data. To bridge this gap, researchers have explored methods to transfer CoT reasoning from LLMs to VLMs. However, existing approaches either need high training costs or require architectural alignment. In this paper, we use Linear Artificial Tomography (LAT) to empirically show that LLMs and VLMs share similar low-frequency latent representations of CoT reasoning despite architectural differences. Based on this insight, we propose L2V-CoT, a novel training-free latent intervention approach that transfers CoT reasoning from LLMs to VLMs. L2V-CoT extracts and resamples low-frequency CoT representations from LLMs in the frequency domain, enabling dimension matching and latent injection into VLMs during inference to enhance reasoning capabilities. Extensive experiments demonstrate that our approach consistently outperforms training-free baselines and even surpasses supervised methods.

Large Language Models (LLMs) excel in reasoning and generation across domains, but still struggle with identifying and diagnosing complex errors. This stems mainly from training objectives that prioritize correct answers, limiting exposure to and learning from errors. While recent studies have begun to address this by introducing error signals, most rely on shallow, static errors, restricting improvement in deep diagnostic ability. To overcome this, we propose Hide and Seek Game (HSG), a dynamic adversarial framework for error generation and diagnosis, and evaluate it on mathematical problem-solving. HSG involves two adversarial roles: Sneaky, which hides by generating subtle, deceptive reasoning errors, and Diagnosis, which seeks to accurately detect them. Through adversarial co-evolution, both error stealth and diagnostic precision are enhanced. Experiments on three mathematical reasoning datasets demonstrate that HSG significantly boosts error diagnosis, achieving 16.8%-31.4% higher accuracy than baselines like GPT-4o. We also release a challenging dataset of deceptive errors and diagnostic annotations as a benchmark for future research.

Diffusion large language models (dLLMs) offer an efficient alternative to autoregressive models through parallel decoding, yet existing post-training methods largely rely on random masking strategies that overlook intrinsic token dependencies.In this work, we present an empirical analysis of attention in dLLMs and show that tokens attending more strongly to unmasked context exhibit greater generation stability and play a critical role in reasoning.Motivated by these findings, we propose AGDO, an attention-guided denoising and optimization framework that aligns both training and optimization with attention-derived dependencies.AGDO determines the denoising order based on attention structure and emphasizes attentioncritical tokens during supervised fine-tuning and reinforcement learning.Experiments on mathematical and coding benchmarks demonstrate that AGDO consistently improves reasoning performance, outperforming state-of-theart post-training methods for dLLMs.

Zhaoheng Huang, Dacheng Wen, Yutao Zhu, Xiaoying Lian, Yushi Liang, Kai Hao, Nan Li, Liangjie Zhang, Qi Zhang, Ji-Rong Wen, Zhicheng Dou, Fangzhao Wu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Retrieval-augmented generation (RAG) effectively enhances the accuracy and timeliness of large language models (LLMs) by incorporating external knowledge retrieved from external sources.However, with the increasing prevalence of LLM-generated content, external corpora used by RAG systems may become contaminated with LLM-generated texts.Such contamination compromises the reliability and quality of retrieved results, ultimately leading to a degradation in RAG performance, and raises concerns about the diminishing presence of human texts and the "Spiral of Silence" effect.A natural solution is to incorporate LLM text detectors into the RAG pipeline to filter out LLM-generated texts from the retrieved results.However, their effective use in RAG remains under-explored.In this paper, we explore the usage paradigms of LLM text detectors for RAG and highlight key limitations of off-the-shelf or directly fine-tuned detectors.To this end, we propose a RAG-aware data augmentation strategy that aligns detector training with realistic contamination patterns.Our approach synthesizes training data from both LLM and human texts under diverse generation modes.Experiments show that our method mitigates performance degradation and improves the long-term stability of RAG systems.

The rapid advancement of large reasoning models has saturated existing math benchmarks, underscoring the urgent need for more challenging evaluation frameworks.To address this, we introduce OlymMATH, a rigorously curated, Olympiad-level math benchmark comprising 350 problems, each with parallel English and Chinese versions.OlymMATH is the first benchmark to unify dual evaluation paradigms within a single suite: (1) natural language evaluation through OlymMATH-EASY and OlymMATH-HARD, comprising 200 computational problems with numerical answers for objective rule-based assessment, and (2) formal verification through OlymMATH-LEAN, offering 150 problems formalized in Lean 4 for rigorous process-level evaluation.All problems are manually sourced from printed publications to minimize data contamination, verified by experts, and span four core domains.Extensive experiments reveal the benchmark's significant challenge, and our analysis also uncovers consistent performance gaps between languages and identifies cases where models employ heuristic "guessing" rather than rigorous reasoning.To further support community research, we release 582k+ reasoning trajectories, a visualization tool, and expert solutions at https://github.com/RUCAIBox/OlymMATH.

Fengqi Zhu, Rongzhen Wang, Shen Nie, Xiaolu Zhang, Chunwei Wu, Jun Zhou, Yankai Lin, Ji-Rong Wen, Chongxuan Li. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Data-driven learning of physical systems has attracted significant attention, where many neural models have been developed. In particular, mesh-based graph neural networks (GNNs) have demonstrated considerable potential in modeling spatiotemporal dynamics across arbitrary geometric domains. However, the existing node-edge message-passing and aggregation mechanism in GNNs limits the representation learning capability. In this paper, we propose a dual-module framework, Cell-embedded and Feature-enhanced Graph Neural Network (CeFeGNN), for learning spatiotemporal dynamics. Specifically, we embed learnable cell attributions to the common node-edge message passing process, thereby better capturing the spatial dependency of regional features. Such a strategy essentially upgrades the local aggregation scheme from first order (e.g., from edge to node) to a higher order (e.g., from volume and edge to node), which takes advantage of volumetric information in message passing. Meanwhile, a novel feature-enhanced block is designed to further improve the model's performance and alleviate the over-smoothing problem. Extensive experiments on various PDE systems and a real-world dataset demonstrate that CeFeGNN achieves superior performance compared with other baselines.

Information access has evolved from search to ask to act. For more than two decades, information retrieval made networked information usable through crawling, indexing, ranking, user modeling, and recommendation. My early work on Web page understanding, vertical search, learning-to-rank, personalized search, and recommendation followed this user-centered view. This keynote briefly revisits that trajectory as the context for a large shift recently: from retrieving documents to constructing systems that reason over evidence, generate grounded answers, and execute information-seeking tasks.

Recently, generative recommendation has emerged as a promising paradigm, attracting significant research attention. The basic framework involves an item tokenizer, which represents each item as a sequence of codes serving as its identifier, and a generative recommender that predicts the next item by autoregressively generating the target item identifier. However, in existing methods, both the tokenizer and the recommender are typically domain-specific, limiting their ability for effective transfer or adaptation to new domains.

Information retrieval (IR) systems have traditionally been designed and trained for human users, with learning-to-rank methods relying heavily on large-scale human interaction logs such as clicks and dwell time. With the rapid emergence of large language model (LLM) powered search agents, however, retrieval is increasingly consumed by agents rather than human beings, and is embedded as a core component within multi-turn reasoning and action loops. In this setting, retrieval models trained under human-centric assumptions can be mismatched with the way agents issue intermediate queries and consume results. In this work, we argue that retrieval models for agentic search should be trained directly from agent interaction data. We study learning to retrieve from agent trajectories as a trajectory-supervised training setting, where supervision is derived from multi-step agent interactions. Through a systematic analysis of search agent trajectories, we identify key behavioral signals that reveal document utility, including browsing actions, unbrowsed rejections, and post-browse reasoning traces. Guided by these insights, we propose LRAT, a simple yet effective framework that mines high-quality retrieval supervision from agent trajectories and incorporates relevance intensity through weighted optimization. To instantiate this setting at scale, we deploy the Tongyi-DeepResearch-30B model on 10K InfoSeekQA queries with four retrievers, collecting 26,482 agent trajectories and constructing 91,713 training pairs. Extensive experiments on both in-domain and out-of-domain deep research benchmarks demonstrate that retrievers trained with LRAT consistently improve evidence recall, end-to-end task success, and execution efficiency across diverse agent architectures and scales. Our results highlight agent trajectories as a practical and scalable supervision source for retrieval in agentic search.

Click-through rate (CTR) prediction plays a pivotal role in online advertising and recommender systems. Despite notable progress in modeling user preferences from historical behaviors, two key challenges persist. First, exsiting discriminative paradigms focus on matching candidates to user history, often overfitting to historically dominant features and failing to adapt to rapid interest shifts. Second, a critical information chasm emerges from the point-wise ranking paradigm. By scoring each candidate in isolation, CTR models discard the rich contextual signal implied by the recalled set as a whole, leading to a misalignment where long-term preferences often override the user's immediate, evolving intent. To address these issues, we propose GenCI, a generative user intent framework that leverages semantic interest cohorts to model dynamic user preferences for CTR prediction. The framework first employs a generative model, trained with a next-item prediction (NTP) objective, to proactively produce candidate interest cohorts. These cohorts serve as explicit, candidate-agnostic representations of a user's immediate intent. A hierarchical candidate-aware network then injects this rich contextual signal into the ranking stage, refining them with cross-attention to align with both user history and the target item. The entire model is trained end-to-end, creating a more aligned and effective CTR prediction pipeline. Extensive experiments on three widely used datasets demonstrate the effectiveness of our approach.

Abstract The rapid evolution of large language models (LLMs) has driven a transformative shift in artificial intelligence (AI), reshaping both research paradigms and practical applications. Distinguished from their predecessors by unprecedented scale and advanced capabilities, LLMs necessitate new frameworks for understanding their development, behavior, and societal impact. This survey systematically reviews recent advancements in LLM techniques across four key dimensions: (1) pre-training methodologies, which establish core model capabilities through large-scale self-supervised training, architectural innovations, and data curation strategies; (2) post-training techniques, including supervised fine-tuning and reinforcement learning, which adapt foundational models to downstream tasks and enhance their alignment and safety; (3) utilization strategies, such as in-context learning, prompt engineering, and agentic reasoning, that optimize real-world deployment and enable effective interaction with external environments; and (4) evaluation methods, encompassing benchmarks for key ability dimensions such as core language capabilities, reasoning, and safety, which support comprehensive and reliable assessment of model performance. Additionally, we identify critical research issues, including those concerning theoretical foundations, efficient scaling, alignment, and agentic capability, and highlight the open challenges they present. By synthesizing state-of-the-art insights and emerging trends, this survey aims to provide a systematic and comprehensive framework for understanding the trajectory, current limitations, and future directions of LLM progress.

Traditional imitation learning focuses on modeling the behavioral mechanisms of experts, which requires a large amount of interaction history generated by some fixed expert. However, in many streaming applications, such as streaming recommender systems, online decision-makers typically engage in online learning during the decision-making process, meaning that the interaction history generated by online decision-makers includes their behavioral evolution from novice expert to experienced expert. This poses a new challenge for existing imitation learning approaches that can only utilize data from experienced experts. To address this issue, this paper proposes an inverse batched contextual bandit (IBCB) framework that can efficiently perform estimations of environment reward parameters and learned policy based on the expert's behavioral evolution history. Specifically, IBCB formulates the inverse problem into a simple quadratic programming problem by utilizing the behavioral evolution history of the batched contextual bandit with inaccessible rewards, and it can be extended to fairness-aware expert limitation. We demonstrate that IBCB is a unified framework for both deterministic and randomized bandit policies. The experimental results indicate that IBCB outperforms several existing imitation learning algorithms on synthetic and real-world data and significantly reduces running time. Additionally, empirical analyses reveal that IBCB exhibits better imitation ability for fairness-aware experts, out-of-distribution generalization and is highly effective in learning the bandit policy from the interaction history of novice experts. The code is publicly available.

Click-through rate (CTR) prediction holds a pivotal role in online advertising and recommendation systems, where even small improvements can significantly boost revenue. Existing research primarily focuses on designing dual-stream architectures to capture effective complex feature interactions from both explicit and implicit perspectives. However, these approaches are faced with two major challenges: (1) the high complexity of feature interaction learning, which increases computational demands and the overfitting risk, and (2) the imbalance between explicit and implicit modules, where one module’s output may dominate the final prediction. To address these issues, in this article, we propose dual-stream multi-layer perceptron (DS-MLP) , a novel feature interaction framework for the CTR prediction task. Specially, it leverages knowledge distillation to consolidate the capacity of learning explicit feature interaction into a main MLP network, while a parallel MLP simultaneously captures implicit feature interactions as a complement. To effectively optimize the DS-MLP architecture, we further design a specific learning approach with two alignment strategies for enhancing the compatibility of the two MLP components. Experiments demonstrate that DS-MLP , though merely a vanilla MLP structure (the final model), can achieve state-of-the-art performance across three widely used benchmarks, offering a scalable and efficient solution for large-scale recommendation systems. Our code is available at https://github.com/RUCAIBox/DS-MLP .

Personalized search has been proven to be an effective method to improve ranking quality by tailoring result lists according to the user's search history. Previous studies achieve personalization by learning a user interest profile from the search log, and decide the candidate document's ranking score by calculating its relevance with the learned profile vector. However, existing approaches overlook fine-grained interaction signals by treating the candidate document separately from the user's search history, relying solely on comparisons with a unified interest vector for re-ranking. Leveraging history-document interactions is not trivial due to the challenge of assessing the contributions of fine-grained matching signals within complex evolving interest patterns. In this paper, we address this challenge by helping the model understand these interactions within the evolving interest process through their integration into the interest profiling procedure. Specifically, we hierarchically incorporate these interaction signals as document-aware interests into behavior representations, employing explicit balancing and differentiation mechanisms, while jointly learning the interest pattern from both actual clues derived from original interests and potential insights provided by document-aware interests. Experimental results show that our model obtains substantial improvements over existing methods.

Search result diversification focuses on providing relevant and diverse documents covering different users’ intents. Intuitively, training an effective and stable search result diversification model needs a large amount of training data. Unfortunately, annotating such data that encompass real users’ search intents is expensive and time-consuming, and most existing models are trained with limited training data, which might lead to unsatisfactory ranking results. Given that Wikipedia contains massive amounts of rigorous editorial and well-structured data, in this article, we propose a pre-training framework leveraging the large-scale Wikipedia data to build weak-supervised signals. Specifically, we introduce four strategies to extract paired supervised signals reflecting the subtopic coverage information from Wikipedia. We also propose a subtopic-disentangled negative sampling strategy to sample hard negative samples and enhance the model’s ability to identify subtle subtopic differences. Four auxiliary tasks are devised to pre-train the Transformer model, which is further adopted as the representation generation model in the downstream diversified ranking. Experimental results demonstrate that our pre-trained model can significantly improve the performance of several existing models, which confirms the effectiveness and scalability of pre-training for search result diversification.

The diversified recommendation aims to satisfy a user’s different preferences and hence alleviates the information cocoon problem. Previous methods focus on increasing the sample probability of interacted items in the long-tail category. However, these methods are limited by the scope of the historical interactions of a single user and confront the challenge of inadequate diversity of past interactions and unpredictable potential diverse preferences. Drawing from social cognitive theory, observational learning ability allows humans to imitate others’ behaviors when their experience is insufficient. Inspired by it, in this article, we apply the idea of observational learning to the diversified recommendation and introduce a social Cog nitive Theory Enhanced D iversified R ecommendation ( Cog4DR ) model. Specifically, we design a three-step observational learning pipeline, including attention, purification, and retention, corresponding to the three essential stages of observational learning. The pipeline enables the current user to observe other users who have similar tastes but also engage with unique categories, therefore exploring potential diverse preferences and achieving dual improvements in accuracy and diversity. Experimental results indicate that Cog4DR outperforms all previous approaches, demonstrating the effectiveness of imitating other users’ behaviors for diversified recommendations.

Recent studies have shown that large language models (LLMs) can assess relevance and support information retrieval (IR) tasks such as document ranking and relevance judgment generation. However, the internal mechanisms by which off-the-shelf LLMs understand and operationalize relevance remain largely unexplored. In this article, we systematically investigate how different LLM modules contribute to relevance judgment through the lens of mechanistic interpretability. Using activation patching techniques, we analyze the roles of various model components and identify a multi-stage, progressive process in generating either pointwise or pairwise relevance judgment. Specifically, LLMs first extract query and document information in the early layers, then process relevance information according to instructions in the middle layers, and finally utilize specific attention heads in the later layers to generate relevance judgments in the required format. Our findings provide insights into the mechanisms underlying relevance assessment in LLMs, offering valuable implications for future research on leveraging LLMs for IR tasks.

This is an erratum for the article “Recommendation as Instruction Following: A Large Language Model Empowered Recommendation Approach” published in ACM Trans. Inf. Syst. 43, 5, Article 114 (July 2025), 37 pages.

As a primary means of information acquisition, information retrieval (IR) systems, such as search engines, have integrated themselves into our daily lives. These systems also serve as components of dialogue, question-answering, and recommender systems. The trajectory of IR has evolved dynamically from its origins in term-based methods to its integration with advanced neural models. While the neural models excel at capturing complex contextual signals and semantic nuances, they still face challenges such as data scarcity, interpretability, and the generation of contextually plausible yet potentially inaccurate responses. This evolution requires a combination of traditional methods (such as term-based sparse retrieval methods with rapid response) and modern neural architectures (such as language models with powerful language understanding capacity). Meanwhile, the emergence of large language models (LLMs) has revolutionized natural language processing due to their remarkable language understanding, generation, and reasoning abilities. Consequently, recent research has sought to leverage LLMs to improve IR systems. Given the rapid evolution of this research trajectory, it is necessary to consolidate existing methodologies and provide nuanced insights through a comprehensive overview. In this survey, we delve into the confluence of LLMs and IR systems, including crucial aspects such as query rewriters, retrievers, rerankers, readers, and search agents.

Retrieval-augmented generation (RAG) is a promising way to improve large language models (LLMs) for generating more factual, accurate, and up-to-date content. Existing methods either optimize prompts to guide LLMs in leveraging retrieved information or directly fine-tune LLMs to adapt to RAG scenarios. Although fine-tuning can yield better performance, it often compromises the LLMs' general generation capabilities by modifying their parameters. This limitation poses challenges in practical applications, especially when LLMs are already deployed, as parameter adjustments may affect their original functionality. To address this, we propose a novel method that involves learning scalable and pluggable virtual tokens for RAG. By maintaining the LLMs' original parameters and fine-tuning only the embeddings of these pluggable tokens, our approach not only enhances LLMs' performance but also preserves their general generation capabilities. Furthermore, we design several training strategies to improve the scalability, flexibility, and generalizability of our method. Comprehensive experiments across 12 question-answering tasks demonstrate the superiority of our approach.

Automatic prompt optimization is an important approach to improving the performance of large language models (LLMs). Recent research demonstrates the potential of using LLMs as prompt optimizers, which can generate improved task prompts via iterative refinement. In this paper, we propose a novel perspective to investigate the design of LLM-based prompt optimizers, by drawing an analogy with gradient-based model optimizers. To connect these two approaches, we identify two pivotal factors in model parameter learning: update direction and update method. By systematically analyzing a rich set of improvement strategies on the two aspects, we further develop a capable Gradient-inspired LLM-based Prompt Optimizer called GPO. At each step, it first retrieves relevant prompts from the optimization trajectory as the update direction. Then, it utilizes the generation-based refinement strategy to perform the update, while controlling the edit distance through a cosine-based decay strategy. Extensive experiments demonstrate the effectiveness and efficiency of GPO. In particular, GPO brings an additional improvement of up to 56.8% on Big-Bench Hard and 62.6% on MMLU compared to baseline methods.

Following natural instructions is crucial for the effective application of Retrieval-Augmented Generation (RAG) systems. Despite recent advancements in Large Language Models (LLMs), research on assessing and improving instruction-following (IF) alignment within the RAG domain remains limited. To address this issue, we propose VIF-RAG, an automated, scalable, and verifiable synthetic pipeline for instruction-following alignment in RAG systems. We start by manually crafting a minimal set of atomic instructions (100k) through automated processes. To further bridge the gap in instruction-following auto-evaluation for RAG systems, we introduce FollowRAG Benchmark, which includes approximately 3K test samples, covering 22 categories of general instruction constraints and four knowledge-intensive QA datasets. Due to its robust pipeline design, FollowRAG can seamlessly integrate with different RAG benchmarks. Using FollowRAG and eight widely-used IF and foundational abilities benchmarks for LLMs, we demonstrate that VIF-RAG markedly enhances LLM performance across a broad range of general instruction constraints while effectively leveraging its capabilities in RAG scenarios. Further analysis offers practical insights for achieving IF alignment in RAG systems.

Continual pre-training (CPT) has been an important approach for adapting language models to specific domains or tasks. In this paper, we comprehensively study its key designs to balance the new abilities while retaining the original abilities, and present an effective CPT method that can greatly improve the Chinese language ability and scientific reasoning ability of LLMs. To achieve it, we design specific data mixture and curriculum strategies based on existing datasets and synthetic high-quality data. Concretely, we synthesize multidisciplinary scientific QA pairs based on related web pages to guarantee the data quality, and also devise the performance tracking and data mixture adjustment strategy to ensure the training stability. For the detailed designs, we conduct preliminary studies on a relatively small model, and summarize the findings to help optimize our CPT method. Extensive experiments on a number of evaluation benchmarks show that our approach can largely improve the performance of Llama-3 (8B), including both the general abilities (+8.81 on C-Eval and +6.31 on CMMLU) and the scientific reasoning abilities (+12.00 on MATH and +4.13 on SciEval). Our model, data, and codes are available at https://github.com/RUC-GSAI/Llama-3-SynE.

Guanting Dong, Chenghao Zhang, Mengjie Deng, Yutao Zhu, Zhicheng Dou, Ji-Rong Wen. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Retrieval-augmented generation (RAG) has emerged as a pivotal technology in natural language processing, owing to its efficacy in generating factual content.However, its informative inputs and complex paradigms often lead to a greater variety of errors.Consequently, achieving automated on-policy assessment and error-oriented correction remains an unresolved issue.In this paper, we propose RAG-Critic, a novel framework that leverages a critic-guided agentic workflow to improve RAG capabilities autonomously.Specifically, we initially design a data-driven error mining pipeline to establish a hierarchical RAG error system.Based on this system, we progressively align an errorcritic model using a coarse-to-fine training objective, which automatically provides finegrained error feedback.Finally, we design a critic-guided agentic RAG workflow that customizes executor-based solution flows based on the error-critic model's feedback, facilitating an error-driven self-correction process.Experimental results across seven RAG-related datasets confirm the effectiveness of RAG-Critic, while qualitative analysis offers practical insights for achieving reliable RAG systems.Our dataset and code are available at https: //github.com/RUC-NLPIR/RAG-Critic.

Due to the immense resource demands and the involved complex techniques, it is still challenging for successfully pre-training a large language models (LLMs) with state-of-the-art performance. In this paper, we explore the key bottlenecks and designs during pre-training, and make the following contributions: (1) a comprehensive investigation into the factors contributing to training instability; (2) a robust optimization approach designed to mitigate training instability effectively; (3) an elaborate data pipeline that integrates data synthesis, data curriculum, and data selection. By integrating the above techniques, we create a rather low-cost training recipe and use it to pre-train YuLan-Mini, a fully-open base model with 2.4B parameters on 1.08T tokens. Remarkably, YuLan-Mini achieves top-tier performance among models of similar parameter scale, with comparable performance to industry-leading models that require significantly more data. To facilitate reproduction, we release the full details of training recipe and data composition. Project details can be accessed at the following link: https://anonymous.4open.science/r/YuLan-Mini/README.md.

In this paper, we aim to improve the reasoning ability of large language models (LLMs) over knowledge graphs (KGs) to answer complex questions.Inspired by existing methods that design the interaction strategy between LLMs and KG, we propose an autonomous LLM-based agent framework, called KG-Agent, which enables a small LLM to actively make decisions until finishing the reasoning process over KGs.In KG-Agent, we integrate the LLM, multifunctional toolbox, KG-based executor, and knowledge memory, and develop an iteration mechanism that autonomously selects the tool and then updates the memory for reasoning over KG.To guarantee the effectiveness, we leverage program language to formulate the multi-hop reasoning process over the KG and synthesize a code-based instruction dataset to fine-tune the base LLM.Extensive experiments demonstrate that only using 10K samples for tuning LLaMA2-7B can outperform competitive methods using larger LLMs or more data, on both in-domain and out-domain datasets.Our code and data will be publicly released.

Recently, large language models (LLMs) have been introduced into recommender systems (RSs) as recommendation backbones or to enhance traditional recommendation models (TRMs). However, existing LLM-based RSs fail to fully leverage the complementary strengths of LLMs (e.g., world knowledge and reasoning capabilities) and TRMs (e.g., recommendation-specific knowledge and computational efficiency), resulting in shallow exploration of the item space. To address this limitation, we propose DeepRec, a novel LLM-based RS approach that facilitates autonomous multi-turn interactions between LLMs and TRMs for deep item space exploration. In each interaction turn, LLMs reason over user preferences and collaborate with TRMs to retrieve candidate items. After multi-turn interaction, LLMs rank the aggregated candidates to generate the final recommendations. We utilize reinforcement learning (RL) for optimization and introduce novel contributions in three key aspects: recommendation model based data rollout, recommendation-oriented hierarchical rewards, and a two-stage RL training strategy. For data rollout, we design a preference-aware TRM, with which LLMs interact to construct trajectory data. For reward design, we propose a hierarchical reward function that comprises both process-level and outcome-level rewards to optimize the interaction process and recommendation quality, respectively. For RL training, our two-stage RL strategy first guides LLMs to learn effective interactions with TRMs, followed by recommendation-oriented RL for performance enhancement. Experiments on public datasets show that DeepRec substantially outperforms both traditional and existing LLM-based baselines, establishing a new paradigm for deep exploration in recommender systems.

Conversational Retrieval-Augmented Generation (RAG) systems enhance user interactions by integrating large language models (LLMs) with external knowledge retrieval. However, multi-turn conversations present significant challenges, including implicit user intent and noisy context, which hinder accurate retrieval and response generation. Existing approaches often struggle with the unstructured conversational context and fail to model explicit relations among conversational turns. Moreover, they do not leverage historically relevant passages effectively. To overcome these limitations, we propose EvoRAG, a novel framework that maintains an evolving knowledge graph aligned with the unstructured conversational context. This graph explicitly captures relations among user queries, system responses, and relevant passages across conversational turns, serving as a structured representation of the context. EvoRAG includes three key components: (1) a dual-path retrieval module for context denoising, (2) a unified knowledge integration module for query rewriting and summarization, and (3) a graph-enhanced RAG module for accurate retrieval and response generation. Experiments on four public conversational RAG datasets show that EvoRAG significantly outperforms strong baselines, particularly in handling topic shifts and long dialogue contexts.

Search result diversification plays a crucial role in addressing query ambiguity and multi-faceted information needs by reducing redundancy across documents. While previous supervised approaches can achieve superior performance, they require costly, large-scale annotated data. In contrast, unsupervised methods are more flexible and training-free but rely on manually designed ranking functions, often leading to suboptimal performance. Inspired by how humans explore diverse information during real-world searching, we propose a diversified search agent DIVAgent to combine the advantages of supervised and unsupervised methods. DIVAgent introduces LLMs as the ''brain'' to reason over complex and diverse search results and delineate human cognitive processes into a workflow tailored for search result diversification. Our search agent first identifies potential user intents and then analyzes the alignment of each document to the intents via an intent-aware module. To guide the generation of diversified document rankings, we design an intent-guided ranker that explicitly links documents to their dominant intents while performing greedy document selection. Experimental results demonstrate that DIVAgent significantly outperforms existing unsupervised baselines and achieves competitive performance with supervised models, highlighting the promise of LLMs for diversified ranking in realistic search scenarios.

Recently, research found that the documents retrieved from the Retrieval-Augmented Generation (RAG) may contain conflicting knowledge with each other, leading Large Language Models (LLMs) to generate incorrect responses. To solve such a problem, existing approaches usually only keep the most frequently mentioned knowledge from these documents, since they assume that the most representative knowledge aligns best with the true answer. Although effective in certain scenarios, these approaches often underperform when the most frequent knowledge is not the correct one. From the voting perspective, these methods can be regarded as a Majority Voting (MV) process, which chooses the most frequent candidates among different candidate knowledge. However, we show that the underperformance of such methods stems from that MV is only effective with a small number of candidates and binary voting scores. In contrast, in the RAG scenario, the candidates (knowledge) are very diverse, and the voting scores (document relevance scores) are typically continuous. Simply adapting MV in RAG will result in poor performance of LLMs. In voting theory, on the other hand, the preference-based voting methods represented by the Borda Voting (BV) consider the whole preference order of voters over all candidates, enabling the selection of candidates that better represent the collective viewpoint. Inspired by such an insight, we propose BordaRAG, a model designed to better select the most appropriate documents from conflicting documents. Specifically, BordaRAG first computes the preference scores of the documents over the candidate answers. After that, a BV component is designed to select the winning documents according to the preference scores. Finally, the chosen documents are provided to LLMs, which will generate the final response. Experimental results on three open-domain QA datasets show that BordaRAG can outperform all baselines.

To satisfy users' diverse information needs, the aggregated search systems need to integrate heterogeneous results, with rich but different structural information, from a variety of verticals, such as news search, video search, and product search. A key challenge in aggregated search is to effectively and efficiently retrieve the most relevant results among a large number of heterogeneous information from different verticals. With the development of deep learning and pre-trained language models (PLMs), many researchers resort to Dense Retrieval (DR) models for a unified, efficient embedding-based retrieval and a better retrieval performance. However, existing dense retrieval models have limitations in: 1) capturing the structural information of search results ; and 2) generalizing across different vertical domains where the search results have different or even unseen structures. In this paper, we aim to tackle these limitations, and propose an effective and efficient dense retrieval model for aggregated search. Specifically, we utilize a deep prompt-tuning technique to make the pre-training model easily applied to downstream vertical search tasks. To capture the structural knowledge, we design a Graph Neural Network (GNN)-based structure prompt, to prompt how text segments are organized in the vanilla semi-structured data. We further incorporate a distributional prompt to model the theme of each domain, and enhance cross-domain generalization. Extensive experiments on the real-world data collected from the WeChat Search demonstrate that for aggregated search tasks, our models can achieve better performance over existing retrieval models, and have the superior ability to generalize to the various or even unseen vertical search tasks.

Unbiased learning to rank (ULTR), which aims to learn unbiased ranking models from biased user behavior logs, plays an important role in Web search. Previous research on ULTR has studied a variety of biases in users' clicks, such as position bias, presentation bias, and outlier bias. However, existing work often assumes that the behavior logs are collected from an ''average'' user, neglecting the differences between different users in their search and browsing behaviors. In this paper, we introduce personalized factors into the ULTR framework, which we term the user-aware ULTR problem. Through a formal causal analysis of this problem, we demonstrate that existing user-oblivious methods are biased when different users have different preferences over queries and personalized propensities of examining documents. To address such a personalized bias, we propose a novel user-aware inverse-propensity-score estimator for learning-to-rank objectives. Specifically, our approach models the distribution of user browsing behaviors for each query and aggregates user-weighted examination probabilities to determine propensities. We theoretically prove that the user-aware estimator is unbiased under some mild assumptions and shows lower variance compared to the straightforward way of calculating a user-dependent propensity for each impression. Finally, we empirically verify the effectiveness of our user-aware estimator by conducting extensive experiments on two semi-synthetic datasets and a real-world dataset.

Visual instruction tuning is crucial for enhancing the zero-shot generalization capability of Multi-modal Large Language Models (MLLMs). In this paper, we aim to investigate a fundamental question: “what makes for good visual instructions”. Through a comprehensive empirical study, we find that instructions focusing on complex visual reasoning tasks are particularly effective in improving the performance of MLLMs, with results correlating to instruction complexity. Based on this insight, we develop a systematic approach to automatically create high-quality complex visual reasoning instructions. Our approach employs a synthesize-complicate-reformulate paradigm, leveraging multiple stages to gradually increase the complexity of the instructions while guaranteeing quality. Based on this approach, we create the ComVint dataset with 32K examples, and fine-tune four MLLMs on it. Experimental results consistently demonstrate the enhanced performance of all compared MLLMs, such as a 27.86% and 27.60% improvement for LLaVA on MME-Perception and MME-Cognition, respectively. Our code and data are publicly available at the link: https://github.com/RUCAIBox/ComVint.

In-Context Learning (ICL) is a critical capability of Large Language Models (LLMs) as it empowers them to comprehend and reason across interconnected inputs. Evaluating the ICL ability of LLMs can enhance their utilization and deepen our understanding of how this ability is acquired at the training stage. However, existing evaluation frameworks primarily focus on language abilities and knowledge, often overlooking the assessment of ICL ability. In this work, we introduce the ICLEval benchmark to evaluate the ICL abilities of LLMs, which encompasses two key sub-abilities: exact copying and rule learning. Through the ICLEval benchmark, we demonstrate that ICL ability is universally present in different LLMs, and model size is not the sole determinant of ICL efficacy. Surprisingly, we observe that ICL abilities, particularly copying, develop early in the pretraining process and stabilize afterward.

Large language models (LLMs) have shown impressive prowess in solving a wide range of tasks with world knowledge. However, it remains unclear how well LLMs are able to perceive their factual knowledge boundaries, particularly under retrieval augmentation settings. In this study, we present the first analysis on the factual knowledge boundaries of LLMs and how retrieval augmentation affects LLMs on open-domain question answering (QA), with a bunch of important findings. Specifically, we focus on three research questions and analyze them by examining QA, priori judgement and posteriori judgement capabilities of LLMs. We show evidence that LLMs possess unwavering confidence in their knowledge and cannot handle the conflict between internal and external knowledge well. Furthermore, retrieval augmentation proves to be an effective approach in enhancing LLMs’ awareness of knowledge boundaries. We further conduct thorough experiments to examine how different factors affect LLMs and propose a simple method to dynamically utilize supporting documents with our judgement strategy. Additionally, we find that the relevance between the supporting documents and the questions significantly impacts LLMs’ QA and judgemental capabilities.

Large language models (LLMs) have shown incredible performance in completing various real-world tasks. The current paradigm of knowledge learning for LLMs is mainly based on learning from examples, in which LLMs learn the internal rule implicitly from a certain number of supervised examples. However, this learning paradigm may not well learn those complicated rules, especially when the training examples are limited. We are inspired that humans can learn the new tasks or knowledge in another way by learning from rules. That is, humans can learn new tasks or grasp new knowledge quickly and generalize well given only a detailed rule and a few optional examples. Therefore, in this paper, we aim to explore the feasibility of this new learning paradigm, which targets on encoding rule-based knowledge into LLMs. We further propose rule distillation, which first uses the strong in-context abilities of LLMs to extract the knowledge from the textual rules, and then explicitly encode the knowledge into the parameters of LLMs by learning from the above in-context signals produced inside the model. Our experiments show that making LLMs learn from rules by our method is much more efficient than example-based learning in both the sample size and generalization ability. Warning: This paper may contain examples with offensive content.

Large reasoning models (LRMs) have exhibited strong performance on complex reasoning tasks, with further gains achievable through increased computational budgets at inference.However, current test-time scaling methods predominantly rely on redundant sampling, ignoring the historical experience utilization, thereby limiting computational efficiency.To overcome this limitation, we propose Sticker-TTS, a novel test-time scaling framework that coordinates three collaborative LRMs to iteratively explore and refine solutions guided by historical attempts.At the core of our framework are distilled key conditions-termed stickers-which drive the extraction, refinement, and reuse of critical information across multiple rounds of reasoning.To further enhance the efficiency and performance of our framework, we introduce a two-stage optimization strategy that combines imitation learning with self-improvement, enabling progressive refinement.Extensive evaluations on three challenging mathematical reasoning benchmarks, including AIME-24, AIME-25, and Olym-MATH, demonstrate that Sticker-TTS consistently surpasses strong baselines, including self-consistency and advanced reinforcement learning approaches, under comparable inference budgets.These results highlight the effectiveness of sticker-guided historical experience utilization.

Multi-lingual ability transfer has become increasingly important for the broad application of large language models (LLMs).Existing work highly relies on training with the multi-lingual ability-related data, which may not be available for low-resource languages.To solve it, we propose a Multilingual Abilities Extraction and Combination approach (MAEC), which decomposes and extracts language-agnostic ability-related weights from LLMs, and combines them across different languages by simple addition and subtraction operations without training.Specifically, our MAEC consists of the extraction and combination stages.In the extraction stage, we firstly locate key neurons that are highly related to specific abilities, and then employ them to extract the transferable ability-related weights.In the combination stage, we further select the ability-related tensors that mitigate the linguistic effects, and design a combining strategy based on them and the languagespecific weights, to build the multi-lingual ability-enhanced LLM.To assess the effectiveness of our approach, we conduct extensive experiments on LLaMA-3 8B on mathematical and scientific tasks in both high-resource and low-resource lingual scenarios.Empirical results have shown that MAEC can effectively and efficiently extract and combine the advanced abilities, achieving comparable performance with PaLM.Resources are available at https://github.com/RUCAIBox/MAET.

Large language models (LLMs) can generate fluent text, raising concerns about misuse in online comments and academic writing, leading to issues like corpus pollution and copyright infringement.Existing LLM text detection methods often rely on features from the logit distribution of the input text.However, the distinction between the LLM-generated and human-written texts may rely on only a few tokens due to the short length or insufficient information in some texts, leading to minimal and hard-to-detect differences in logit distributions.To address this, we propose HALO, an LLM-based detection method that leverages external text corpora to evaluate the difference in the logit distribution of input text under retrieved human-written and LLM-rewritten contexts.HALO also complements basic detection features and can serve as a plug-and-play module to enhance existing detection methods.Extensive experiments on five public datasets with three widely-used source LLMs show that our proposed detection method achieves state-ofthe-art performance in AUROC, both in crossdomain and domain-specific scenarios.

retrieval-augmented large language models via transferable adversarial attacks.

Existing studies have optimized retrievalaugmented generation (RAG) across various sub-tasks, such as query understanding and retrieval refinement, but integrating these optimizations into a unified framework remains challenging.To tackle this problem, this work proposes RoleRAG, a unified RAG framework that achieves efficient multi-task processing through role-specific token optimization.RoleRAG comprises six modules, each handling a specific sub-task within the RAG process.Additionally, we introduce a query graph to represent the decomposition of the query, which can be dynamically resolved according to the decomposing state.All modules are driven by the same underlying LLM, distinguished by task-specific role tokens that are individually optimized.This design allows RoleRAG to dynamically activate different modules within a single LLM instance, thereby streamlining deployment and reducing resource consumption.Experimental results on five open-domain question-answering datasets demonstrate the effectiveness, generalizability, and flexibility of our framework.

The goal of transferable sequential recommendation (TSR) is to improve the performance of sequential recommenders in multiple target domains leveraging knowledge transferred from source domains. Most existing transferable sequential recommenders rely on item modality information but pay insufficient attention to the significant differences across domains (i.e., distinctive user behavior characteristics) in recommender systems. Besides, they are typically built on specific pre-trained model structures, limiting their model generalization to various different downstream domains and sequential recommenders in practical applications. This paper aims to address these issues by exploring the potential of mixture-of-LoRAs based domain adaption and an ensemble of sequential recommenders for learning generalizable and transferable domain-adaptive item representations. We propose MoL-TSR, a pre-trained plug-in mixture of low-rank adaption model for TSR. Specifically, we first devise a mixture-ofLoRAs based domain projector to produce domain-adaptive item representations based on both item modality and user behavior information for recommendation, enhancing the item discrimination and domain transfer ability. Then, we pre-train an ensemble of sequential recommendation models with contrastive learning to enhance the model-aspect generalization. As a general pretrained model, the mixture-of-LoRAs based domain projector in our MoL-TSR is a plug-in item semantic representation learning module, which is convenient to migrate into other transferable sequential recommenders to enhance item semantics better suited for recommendation. Extensive experiments demonstrate the effectiveness and model generalization of MoL-TSR, promising a practical solution for learning generalizable and transferable item semantic representations in recommendation. Our code is available at https://github.com/RUCAIBox/MoL-TSR.

Video Multimodal Large Language Models~(MLLMs) have shown remarkable capability of understanding the video semantics on various downstream tasks. Despite the advancements, there is still a lack of systematic research on visual context representation, which refers to the scheme to select frames from a video and further select the tokens from a frame. In this paper, we explore the design space for visual context representation, and aim to improve the performance of video MLLMs by finding more effective representation schemes. Firstly, we formulate the task of visual context representation as a constrained optimization problem, and model the language modeling loss as a function of the number of frames and the number of embeddings (or tokens) per frame, given the maximum visual context window size. Then, we explore the scaling effects in frame selection and token selection respectively, and fit the corresponding function curve by conducting extensive empirical experiments. We examine the effectiveness of typical selection strategies and present empirical findings to determine the two factors. Furthermore, we study the joint effect of frame selection and token selection, and derive the optimal formula for determining the two factors. We demonstrate that the derived optimal settings show alignment with the best-performed results of empirical experiments. The data and code are available at: https://github.com/RUCAIBox/Opt-Visor.

The emergence of in-context learning (ICL) is potentially attributed to two major abilities: task recognition (TR) for recognizing the task from demonstrations and utilizing pre-trained priors, and task learning (TL) for learning from demonstrations. However, relationships between the two abilities and how such relationships affect the emergence of ICL is unclear. In this paper, we take the first step by examining the pre-training dynamics of the emergence of ICL. With carefully designed metrics, we find that these two abilities are, in fact, competitive during pre-training. Moreover, we observe a negative correlation between the competition and the performance of ICL. Further analysis of common pre-training factors (i.e., model size, dataset size, and data curriculum) demonstrates possible ways to regulate the competition. Based on these insights, we propose a simple yet effective method to better integrate these two abilities for ICL at inference time. Through adaptive ensemble learning, the performance of ICL can be significantly boosted, enabling two small models to outperform a larger one with more than twice the parameters.

$\textit{Cis}$-regulatory elements (CREs), such as promoters and enhancers, are relatively short DNA sequences that directly regulate gene expression. The fitness of CREs, measured by their ability to modulate gene expression, highly depends on the nucleotide sequences, especially specific motifs known as transcription factor binding sites (TFBSs). Designing high-fitness CREs is crucial for therapeutic and bioengineering applications. Current CRE design methods are limited by two major drawbacks: (1) they typically rely on iterative optimization strategies that modify existing sequences and are prone to local optima, and (2) they lack the guidance of biological prior knowledge in sequence optimization. In this paper, we address these limitations by proposing a generative approach that leverages reinforcement learning (RL) to fine-tune a pre-trained autoregressive (AR) model. Our method incorporates data-driven biological priors by deriving computational inference-based rewards that simulate the addition of activator TFBSs and removal of repressor TFBSs, which are then integrated into the RL process. We evaluate our method on promoter design tasks in two yeast media conditions and enhancer design tasks for three human cell types, demonstrating its ability to generate high-fitness CREs while maintaining sequence diversity. The code is available at https://github.com/yangzhao1230/TACO.

Large language models (LLMs) have become increasingly proficient at simulating various personality traits, an important capability for supporting related applications (e.g., role-playing). To further improve this capacity, in this paper, we present a neuron based approach for personality trait induction in LLMs, with three major technical contributions. First, we construct PERSONALITYBENCH, a large-scale dataset for identifying and evaluating personality traits in LLMs. This dataset is grounded in the Big Five personality traits from psychology and designed to assess the generative capabilities of LLMs towards specific personality traits. Second, by leveraging PERSONALITYBENCH, we propose an efficient method for identifying personality-related neurons within LLMs by examining the opposite aspects of a given trait. Third, we develop a simple yet effective induction method that manipulates the values of these identified personality-related neurons, which enables fine-grained control over the traits exhibited by LLMs without training and modifying model parameters. Extensive experiments validates the efficacy of our neuron identification and trait induction methods. Notably, our approach achieves comparable performance as fine-tuned models, offering a more efficient and flexible solution for personality trait induction in LLMs.

Adapting large language models (LLMs) to specialized domains typically requires domain-specific corpora for continual pre-training to facilitate knowledge memorization and related instructions for fine-tuning to apply this knowledge. However, this method may lead to inefficient knowledge memorization due to a lack of awareness of knowledge utilization during the continual pre-training and demands LLMs to simultaneously learn knowledge utilization and format alignment with divergent training objectives during the fine-tuning. To enhance the domain adaptation of LLMs, we revise this process and propose a new domain adaptation framework including domain knowledge learning and general format alignment, called \emph{Mix-CPT}. Specifically, we first conduct a knowledge mixture continual pre-training that concurrently focuses on knowledge memorization and utilization. To avoid catastrophic forgetting, we further propose a logit swap self-distillation constraint. By leveraging the knowledge and capabilities acquired during continual pre-training, we then efficiently perform instruction tuning and alignment with a few general training samples to achieve format alignment. Extensive experiments show that our proposed \emph{Mix-CPT} framework can simultaneously improve the task-solving capabilities of LLMs on the target and general domains.

Tool learning enables Large Language Models (LLMs) to interact with external environments by invoking tools, serving as an effective strategy to mitigate the limitations inherent in their pre-training data. In this process, tool documentation plays a crucial role by providing usage instructions for LLMs, thereby facilitating effective tool utilization. This paper concentrates on the critical challenge of bridging the comprehension gap between LLMs and external tools due to the inadequacies and inaccuracies inherent in existing human-centric tool documentation. We propose a novel framework, DRAFT, aimed at Dynamically Refining tool documentation through the Analysis of Feedback and Trials emanating from LLMs' interactions with external tools. This methodology pivots on an innovative trial-and-error approach, consisting of three distinct learning phases: experience gathering, learning from experience, and documentation rewriting, to iteratively enhance the tool documentation. This process is further optimized by implementing a diversity-promoting exploration strategy to ensure explorative diversity and a tool-adaptive termination mechanism to prevent overfitting while enhancing efficiency. Extensive experiments on multiple datasets demonstrate that DRAFT's iterative, feedback-based refinement significantly ameliorates documentation quality, fostering a deeper comprehension and more effective utilization of tools by LLMs. Notably, our analysis reveals that the tool documentation refined via our approach demonstrates robust cross-model generalization capabilities.

Reinforcement Learning from Human Feedback (RLHF) has emerged as a standard and effective approach for training large language models (LLMs) with human preferences. In this framework, a learned reward model approximates human preferences and guides policy optimization, making it crucial to develop an accurate reward model. However, without the ``true'' reward function, challenges arise when the reward model is an imperfect proxy for human preference. Since the policy optimization continuously shifts the human preference training dataset's distribution. The fixed reward model suffers from this problem of off-distribution, especially the on policy methods. While collecting new preference data can mitigate this issue, it is costly and challenging to optimize. Thus, reusing the policy interaction samples becomes a possible way to further refine the reward model. To tackle these challenges, we introduce a novel method \textbf{U}ncertainty-\textbf{G}radient based \textbf{D}ata \textbf{A}ugmentation (\textbf{UGDA} for short) to enhance reward modeling by leveraging policy samples to maintain on-distribution performance. Specifically, UGDA selects interaction samples based on the uncertainty of the reward ensembles and the gradient based influence of policy optimization. After the reward relabeling of selected samples, we use supervised learning to refine the reward ensembles, then get the retrained policy. Extensive experiments demonstrate that by leveraging UGDA to select a few samples without the costly human preference data collection, we can improve the ability of the policy and surpass the state-of-the-art methods.

Previous studies have found that PLM-based retrieval models exhibit a preference for LLM-generated content, assigning higher relevance scores to these documents even when their semantic quality is comparable to human-written ones. This phenomenon, known as source bias, threatens the sustainable development of the information access ecosystem. However, the underlying causes of source bias remain unexplored. In this paper, we explain the process of information retrieval with a causal graph and discover that PLM-based retrievers learn perplexity features for relevance estimation, causing source bias by ranking the documents with low perplexity higher. Theoretical analysis further reveals that the phenomenon stems from the positive correlation between the gradients of the loss functions in language modeling task and retrieval task. Based on the analysis, a causal-inspired inference-time debiasing method is proposed, called **C**ausal **D**iagnosis and **C**orrection (CDC). CDC first diagnoses the bias effect of the perplexity and then separates the bias effect from the overall estimated relevance score. Experimental results across three domains demonstrate the superior debiasing effectiveness of CDC, emphasizing the validity of our proposed explanatory framework. Source codes are available at https://github.com/WhyDwelledOnAi/Perplexity-Trap.

Superalignment, where humans act as weak supervisors for superhuman models, has become a crucial problem with the rapid development of Large Language Models (LLMs). Recent work has preliminarily studied this problem by using weak models to supervise strong models, and discovered that weakly supervised strong students can consistently outperform weak teachers towards the alignment target, leading to a weak-to-strong generalization phenomenon. However, we are concerned that behind such a promising phenomenon, whether there exists an issue of weak-to-strong deception, where strong models deceive weak models by exhibiting well-aligned in areas known to weak models but producing misaligned behaviors in cases weak models do not know. We take an initial step towards exploring this security issue in a specific but realistic multi-objective alignment case, where there may be some alignment targets conflicting with each other (e.g., helpfulness v.s. harmlessness). We aim to explore whether, in such cases, strong models might deliberately make mistakes in areas known to them but unknown to weak models within one alignment dimension, in exchange for a higher reward in another dimension. Through extensive experiments in both the reward modeling and preference optimization scenarios, we find: (1) The weak-to-strong deception phenomenon exists across all settings. (2) The deception intensifies as the capability gap between weak and strong models increases. (3) Bootstrapping with an intermediate model can mitigate the deception to some extent, though its effectiveness remains limited. Our work highlights the urgent need to pay more attention to the true reliability of superalignment.

Solving partial differential equations (PDEs) by numerical methods meet computational cost challenge for getting the accurate solution since fine grids and small time steps are required. Machine learning can accelerate this process, but struggle with weak generalizability, interpretability, and data dependency, as well as suffer in long-term prediction. To this end, we propose a PDE-embedded network with multiscale time stepping (MultiPDENet), which fuses the scheme of numerical methods and machine learning, for accelerated simulation of flows. In particular, we design a convolutional filter based on the structure of finite difference stencils with a small number of parameters to optimize, which estimates the equivalent form of spatial derivative on a coarse grid to minimize the equation's residual. A Physics Block with a 4th-order Runge-Kutta integrator at the fine time scale is established that embeds the structure of PDEs to guide the prediction. To alleviate the curse of temporal error accumulation in long-term prediction, we introduce a multiscale time integration approach, where a neural network is used to correct the prediction error at a coarse time scale. Experiments across various PDE systems, including the Navier-Stokes equations, demonstrate that MultiPDENet can accurately predict long-term spatiotemporal dynamics, even given small and incomplete training data, e.g., spatiotemporally down-sampled datasets. MultiPDENet achieves the state-of-the-art performance compared with other neural baseline models, also with clear speedup compared to classical numerical methods

In recent years, substantial research efforts have been devoted to enhancing sequential recommender systems by integrating abundant side information with ID-based collaborative information. This study specifically focuses on leveraging the textual metadata (e.g., titles and brands) associated with items. While existing methods have achieved notable success by combining text and ID representations, they often struggle to strike a balance between textual information embedded in text representations and collaborative information from sequential patterns of user behavior. In light of this, we propose CCFRec, a novel Code-based textual and Collaborative semantic Fusion method for sequential Recommendation. The key idea behind our approach is to bridge the gap between textual and collaborative information using semantic codes. Specifically, we generate fine-grained semantic codes from multi-view text embeddings through vector quantization techniques. Subsequently, we develop a code-guided semantic-fusion module based on the cross-attention mechanism to flexibly extract and integrate relevant information from text representations. In order to further enhance the fusion of textual and collaborative semantics, we introduce an optimization strategy that employs code masking with two specific objectives: masked code modeling and masked sequence alignment. The merit of these objectives lies in leveraging mask prediction tasks and augmented item representations to capture code correlations within individual items and enhance the sequence modeling of the recommendation backbone. Extensive experiments conducted on four public datasets demonstrate the superiority of CCFRec, showing significant improvements over various sequential recommendation models. Our code is available at https://github.com/RUCAIBox/CCFRec.

Recent advances in retrieval-augmented generation (RAG) furnish large language models (LLMs) with iterative retrievals of relevant information to handle complex multi-hop questions. These methods typically alternate between LLM reasoning and retrieval to accumulate external information into the LLM's context. However, the ever-growing context inherently imposes an increasing burden on the LLM to perceive connections among critical information pieces, with futile reasoning steps further exacerbating this overload issue. In this paper, we present KnowTrace, an elegant RAG framework to (1) mitigate the context overload and (2) bootstrap higher-quality multi-step reasoning. Instead of simply piling the retrieved contents, KnowTrace autonomously traces out desired knowledge triplets to organize a specific knowledge graph relevant to the input question. Such a structured workflow not only empowers the LLM with an intelligible context for inference, but also naturally inspires a reflective mechanism of knowledge backtracing to identify contributive LLM generations as process supervision data for self-bootstrapping. Extensive experiments show that KnowTrace consistently surpasses existing methods across three multi-hop question answering benchmarks, and the bootstrapped version further amplifies the gains.

Data-centric methods have shown great potential in understanding and predicting spatiotemporal dynamics, enabling better design and control of the object system. However, deep learning models often lack interpretability, fail to obey intrinsic physics, and struggle to cope with the various domains. While geometry-based methods, e.g., graph neural networks (GNNs), have been proposed to further tackle these challenges, they still need to find the implicit physical laws from large datasets and rely excessively on rich labeled data. In this paper, we herein introduce the conservation-informed GNN (CiGNN), an end-to-end explainable learning framework, to learn spatiotemporal dynamics based on limited training data. The network is designed to conform to the general conservation law via symmetry, where conservative and non-conservative information passes over a multiscale space enhanced by a latent temporal marching strategy. The efficacy of our model has been verified in various spatiotemporal systems based on synthetic and real-world datasets, showing superiority over baseline models. Results demonstrate that CiGNN exhibits remarkable accuracy and generalizability, and is readily applicable to learning for prediction of various spatiotemporal dynamics in a spatial domain with complex geometry.

Visual instruction tuning is the key to building large vision language models (LVLMs), which can greatly improve the task solving and generalization capabilities. Previous work mostly collects a mixture of existing visual instruction datasets via heuristic ways for train- ing (even more than a million instructions), which may introduce data redundancy and increase the training cost. To investigate it, we conduct a series of empirical studies, which show that greatly reducing the amount of instructions from several tasks even do not affect the performance, indicating significant redundancy within the visual instruction datasets. Based on the findings, we propose a high-value data selection approach TIVE, to eliminate redundancy within the visual instruction data and reduce the training cost. In TIVE, based on the gradient-based influence functions, we estimate the instance influence score on its corresponding task and the task difficulty score. Then, we leverage these scores to determine the task proportion within the visual instruction subset, and select high-value instances for each task, respectively. Experiments on various LVLMs show that our approach using only about 15% data can achieve comparable performance to the full-data fine-tuned model across eight benchmarks, even surpassing it on four of the benchmarks. Our code and data will be publicly released.

Xueyang Feng, Bo Lan, Quanyu Dai, Lei Wang, Jiakai Tang, Xu Chen, Zhenhua Dong, Ji-Rong Wen. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Xinyu Tang, Xiaolei Wang, Xin Zhao, Ji-Rong Wen. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Lei Wang, Jianxun Lian, Yi Huang, Yanqi Dai, Haoxuan Li, Xu Chen, Xing Xie, Ji-Rong Wen. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

We present a systematic theoretical framework that interprets masked diffusion models (MDMs) as solutions to energy minimization problems in discrete optimal transport. Specifically, we prove that three distinct energy formulations—kinetic, conditional kinetic, and geodesic energy—are mathematically equivalent under the structure of MDMs, and that MDMs minimize all three when the mask schedule satisfies a closed-form optimality condition. This unification not only clarifies the theoretical foundations of MDMs, but also motivates practical improvements in sampling. By parameterizing interpolation schedules via Beta distributions, we reduce the schedule design space to a tractable 2D search, enabling efficient post-training tuning without model modification. Experiments on synthetic and real-world benchmarks demonstrate that our energy-inspired schedules outperform hand-crafted baselines, particularly in low-step sampling settings.

Generating flexible-view 3D scenes, including 360° rotation and zooming, from single images is challenging due to a lack of 3D data. To this end, we introduce FlexWorld, a novel framework that progressively constructs a persistent 3D Gaussian splatting representation by synthesizing and integrating new 3D content. To handle novel view synthesis under large camera variations, we leverage an advanced pre-trained video model fine-tuned on accurate depth-estimated training pairs. By combining geometry-aware scene integration and optimization, FlexWorld refines the scene representation, producing visually consistent 3D scenes with flexible viewpoints. Extensive experiments demonstrate the effectiveness of FlexWorld in generating high-quality novel view videos and flexible-view 3D scenes from single images, achieving superior visual quality under multiple popular metrics and datasets compared to existing state-of-the-art methods. Additionally, FlexWorld supports extrapolating from existing 3D scenes, further extending its applicability. Qualitatively, we highlight that FlexWorld can generate high-fidelity scenes that enable 360° rotations and zooming exploration. Our code is available at https://github.com/ML-GSAI/FlexWorld.

Large reasoning models (LRMs), such as OpenAI-o1 and DeepSeek-R1, demonstrate impressive long-horizon reasoning capabilities. However, their reliance on static internal knowledge limits their performance on complex, knowledge-intensive tasks and hinders their ability to produce comprehensive research reports requiring synthesis of diverse web information. To address this, we propose WebThinker, a deep research agent that empowers LRMs to autonomously search the web, navigate among web pages, and draft reports during the reasoning process. WebThinker integrates a Deep Web Explorer module, enabling LRMs to dynamically search, navigate, and extract information from the web when encountering knowledge gaps. It also employs an Autonomous Think-Search-and-Draft strategy, allowing the model to seamlessly interleave reasoning, information gathering, and report writing in real time. To further enhance research tool utilization, we introduce an RL-based training strategy via iterative online Direct Preference Optimization (DPO). Extensive experiments on complex reasoning benchmarks (GPQA, GAIA, WebWalkerQA, HLE) and scientific report generation tasks (Glaive) demonstrate that WebThinker significantly outperforms existing methods and strong proprietary systems. Our approach enhances LRM reliability and applicability in complex scenarios, paving the way for more capable and versatile deep research systems. The code is available at https://github.com/RUC-NLPIR/WebThinker.

The capabilities of large language models (LLMs) are widely regarded as relying on autoregressive models (ARMs). We challenge this notion by introducing *LLaDA*, a diffusion model trained from scratch under the pre-training and supervised fine-tuning (SFT) paradigm. LLaDA employs a forward data masking process and a reverse generation process, parameterized by a Transformer to predict masked tokens. It provides a principled generative approach for probabilistic inference by optimizing a likelihood lower bound. Across extensive benchmarks on general tasks, math, code, and so on, LLaDA demonstrates strong *scalability* and performs comparably to our self-constructed ARM baselines. Remarkably, LLaDA 8B is competitive with strong LLMs like LLaMA3 8B in *in-context learning* and, after SFT, exhibits impressive *instruction-following* abilities in case studies such as multi-turn dialogue. Moreover, LLaDA addresses the reversal curse, surpassing GPT-4o in a reversal poem completion task. Our findings show the promise of diffusion models for language modeling at scale and challenge the common assumption that core LLM capabilities discussed above inherently depend on ARMs. Project page and codes: \url{https://ml-gsai.github.io/LLaDA-demo/}.

In this paper, we propose a novel optimizer state compression algorithm, namely \textbf{$\pi$-Quant}, which leverages the properties of irrational numbers (\eg $\pi$) for memory-efficient training. The core idea is based on our mathematical findings, which show that a pair of parameters can be represented by a single rotation angle using the complex rotation scheme. Building on this insight, we map the parameters into a complex space and perform quantization using the corresponding rotation angles. To efficiently integrate it into optimization process, we develop an efficient system of geometric equations that computes the precise rotation angles with linear complexity. We evaluate $\pi$-Quant on a wide range of tasks. Our experiments show that it can reduce the bit-width of parameters to 3.32-bit, achieving a 41.8\% decrease in GPU memory usage, all while maintaining full accuracy. \textcolor{blue}{We have submitted the code in supplementary materials}.

With the significant progress of large reasoning models in complex coding and reasoning tasks, existing benchmarks, like LiveCodeBench and CodeElo, are insufficient to evaluate the coding capabilities of large language models (LLMs) in real competition environments. Moreover, current evaluation metrics such as Pass@K fail to capture the reflective abilities of reasoning models. To address these challenges, we propose ICPC-Eval, a top-level competitive coding benchmark designed to probing the frontiers of LLM reasoning. ICPC-Eval includes 118 carefully curated problems from 11 recent ICPC contests held in various regions of the world, offering three key contributions: 1) A challenging realistic ICPC competition scenario, featuring a problem type and difficulty distribution consistent with actual contests. 2) A robust test case generation method and a corresponding local evaluation toolkit, enabling efficient and accurate local evaluation. 3) An effective test-time scaling evaluation metric, Refine@K, which allows iterative repair of solutions based on execution feedback. The results underscore the significant challenge in evaluating complex reasoning abilities: top-tier reasoning models like DeepSeek-R1 often rely on multi-turn code feedback to fully unlock their in-context reasoning potential when compared to non-reasoning counterparts. Furthermore, despite recent advancements in code generation, these models still lag behind top-performing human teams. We release the benchmark at: https://github.com/RUCAIBox/ICPC-Eval

Accurately locating key moments within long videos is crucial for solving long video understanding (LVU) tasks. However, existing benchmarks are either severely limited in terms of video length and task diversity, or they focus solely on the end-to-end LVU performance, making them inappropriate for evaluating whether key moments can be accurately accessed. To address this challenge, we propose MomentSeeker, a novel benchmark for long-video moment retrieval (LVMR), distinguished by the following features. First, it is created based on long and diverse videos, averaging over 1,200 seconds in duration, and collected from various domains, e.g., movie, anomaly, egocentric, and sports. Second, it covers a variety of real-world scenarios in three levels: global-level, event-level, and object-level, covering common tasks like action recognition, object localization, causal reasoning, etc. Third, it incorporates rich forms of queries, including text-only queries, image-conditioned queries, and video-conditioned queries. On top of MomentSeeker, we conduct comprehensive experiments for both generation-based approaches (directly using MLLMs) and retrieval-based approaches (leveraging video retrievers). Our results reveal the significant challenges in long-video moment retrieval in terms of accuracy and efficiency, despite improvements from the latest long-video MLLMs and task-specific fine-tuning. We have publicly released MomentSeeker to facilitate future research in this area.

LLM-based agents have been widely applied as personal assistants, capable of memorizing information from user messages and responding to personal queries. However, there still lacks an objective and automatic evaluation on their memory capability, largely due to the challenges in constructing reliable questions and answers (QAs) according to user messages. In this paper, we propose MemSim, a Bayesian simulator designed to automatically construct reliable QAs from generated user messages, simultaneously keeping their diversity and scalability. Specifically, we introduce the Bayesian Relation Network (BRNet) and a causal generation mechanism to mitigate the impact of LLM hallucinations on factual information, facilitating the automatic creation of an evaluation dataset. Based on MemSim, we generate a dataset in the daily-life scenario, named MemDaily, and conduct extensive experiments to assess the effectiveness of our approach. We also provide a benchmark for evaluating different memory mechanisms in LLM-based agents with the MemDaily dataset.

Recommender systems are essential for information access, allowing users to present their content for recommendation. With the rise of large language models (LLMs), AI-generated content (AIGC), primarily in the form of text, has become a central part of the content ecosystem. As AIGC becomes increasingly prevalent, it is important to understand how it affects the performance and dynamics of recommender systems. To this end, we construct an environment that incorporates AIGC to explore its short-term impact. The results from popular sequential recommendation models reveal that AIGC are ranked higher in the recommender system, reflecting the phenomenon of source. To further explore the long-term impact of AIGC, we introduce a feedback loop with realistic simulators. The results show that the model's preference for AIGC increases as the user clicks on AIGC rises and the model trains on simulated click data. This leads to two issues: In the short term, bias toward AIGC encourages LLM-based content creation, increasing AIGC content, and causing unfair traffic distribution. From a long-term perspective, our experiments also show that when AIGC dominates the content ecosystem after a feedback loop, it can lead to a decline in recommendation performance. To address these issues, we propose a debiasing method based on L1-loss optimization to maintain long-term content ecosystem balance. In a real-world environment with AIGC generated by mainstream LLMs, our method ensures a balance between AIGC and human-generated content in the ecosystem. The code and dataset are available at https://github.com/Yuqi-Zhou/Rec_SourceBias.

Conversational recommendation systems (CRSs) use multi-turn interaction to capture user preferences and provide personalized recommendations. A fundamental challenge in CRSs lies in effectively understanding user preferences from conversations. Previous research primarily focuses on the issue of insufficient contextual information in conversations. They address this by introducing external knowledge sources, such as knowledge graphs, large language models (LLMs), and conversational recommendation corpora. Based on this, they design specific alignment strategies (e.g., prompt learning and instruction tuning) to integrate such knowledge for user preference understanding and item recommendation. However, user preferences can be multifaceted and complex, posing significant challenges for accurate recommendations even with access to abundant external knowledge. While interaction with users can clarify their true preferences, frequent user involvement may lead to a degraded user experience.

In modern information retrieval (IR), going beyond accuracy is crucial for maintaining a healthy ecosystem, particularly in meeting fairness and diversity requirements. To address these needs, various datasets, algorithms, and evaluation methods have been developed. These algorithms are often tested with different metrics, datasets, and experimental settings, making comparisons inconsistent and challenging. Consequently, there is an urgent need for a comprehensive IR toolkit, enabling standardized assessments of fairness- and diversity-aware algorithms across IR tasks. To address these issues, we introduce an open-source standardized toolkit called FairDiverse. First, FairDiverse provides a comprehensive framework for incorporating fairness- and diversity-aware approaches, including pre-processing, in-processing, and post-processing methods, into different pipeline stages of IR. Second, FairDiverse enables the evaluation of 29 fairness, and diversity algorithms across 16 base models for two fundamental IR tasks-search and recommendation-facilitating the establishment of a comprehensive benchmark. Finally, FairDiverse is highly extensible, offering multiple APIs to enable IR researchers to quickly develop their own fairness- and diversity-aware IR models, and allows for fair comparisons with existing baselines. The project is open-sourced on GitHub:~ https://github.com/XuChen0427/FairDiverse.

Generative AI search driven by large language models (LLMs) is reshaping information retrieval by offering end-to-end answers to complex queries, reducing users' reliance on manually browsing and summarizing multiple web pages. However, while this paradigm enhances convenience, it disrupts the feedback-driven improvement loop that has historically powered the evolution of traditional Web search. Web search can continuously improve their ranking models by collecting large-scale, fine-grained user feedback (e.g., clicks, dwell time) at the document level. In contrast, generative AI search operates through a much longer search pipeline-spanning query decomposition, document retrieval, and answer generation-yet typically receives only coarse-grained feedback on the final answer. This introduces a feedback loop disconnect, where user feedback for the final output cannot be effectively mapped back to specific system components, making it difficult to improve each intermediate stage and sustain the feedback loop.

Counterfactual learning to rank (CLTR) has attracted extensive attention in the IR community for its ability to leverage massive logged user interaction data to train ranking models. While the CLTR models can be theoretically unbiased when the user behavior assumption is correct and the propensity estimation is accurate, their effectiveness is usually empirically evaluated via simulation-based experiments due to a lack of widely available, large-scale, real click logs. However, many previous simulation-based experiments are somewhat limited because they may have one or more of the following deficiencies: 1) using a weak production ranker to generate initial ranked lists, 2) relying on a simplified user simulation model to simulate user clicks, and 3) generating a fixed number of synthetic click logs. As a result, the robustness of CLTR models in complex and diverse situations is largely unknown and needs further investigation.

Maintaining the long-term sustainability of recommender systems (RS) is crucial.Traditional RS evaluation methods primarily focus on the user's immediate feedback (e.g., click), however, they often overlook the long-term effect involved by the content creators.In the real world, content creators can strategically create and upload new items to the platform by analyzing users' feedback and preference trends.Although previous studies have attempted to model creator behaviors, they often overlook that such behaviors are under conditions of information asymmetry.This asymmetry arises because creators mainly access the user feedback on the items they produce, while the platform has access to the full spectrum of feedback data.However, existing RS simulators often fail to consider such a condition, making the long-term RS evaluation inaccurate.To bridge this gap, we propose a Large Language Model (LLM)empowered creator simulation agent named CreAgent.By utilizing the belief mechanism from game theory and the fast-and-slow thinking framework, we can simulate the creator's behaviors well under information asymmetry.Furthermore, to enhance CreAgent's simulation ability, we utilize Proximal Policy Optimization to fine-tune CreAgent.Our credibility validation experiments demonstrate that our simulation environment effectively aligns with the behaviors of real-world platforms and creators, thereby enhancing the reliability of long-term evaluations in RS.Furthermore, leveraging this simulator, we can examine whether RS algorithms, such as fairnessand diversity-aware methods, contribute to improving long-term performance for different stakeholders.

Text-guided image generation and editing using diffusion models have achieved remarkable advancements. Among these, tuning-free methods have gained attention for their ability to perform edits without extensive model adjustments, offering simplicity and efficiency. However, existing tuning-free approaches often struggle with balancing fidelity and editing precision. Reconstruction errors in DDIM Inversion are partly attributed to the cross-attention mechanism in U-Net, which introduces misalignments during the inversion and reconstruction process. To address this, we analyze reconstruction from a structural perspective and propose a novel approach that replaces traditional cross-attention with uniform attention maps, significantly enhancing image reconstruction fidelity. Our method effectively minimizes distortions caused by varying text conditions during noise prediction. To complement this improvement, we introduce an adaptive mask-guided editing technique that integrates seamlessly with our reconstruction approach, ensuring consistency and accuracy in editing tasks. Experimental results demonstrate that our approach not only excels in achieving high-fidelity image reconstruction but also performs robustly in real image composition and editing scenarios. This study underscores the potential of uniform attention maps to enhance the fidelity and versatility of diffusion-based image processing methods. Code is available at https://github.com/Mowenyii/Uniform-Attention-Maps.

Retrieval-augmented generation (RAG) has effectively mitigated the hallucination problem of large language models (LLMs). However, the difficulty of aligning the retriever with the LLMs' diverse knowledge preferences inevitably poses a challenge in developing a reliable RAG system. To address this issue, we propose DPA-RAG, a universal framework designed to align diverse knowledge preferences within RAG systems. Specifically, we initially introduce a preference knowledge construction pipeline and incorporate five novel query augmentation strategies to alleviate preference data scarcity. Based on preference data, DPA-RAG accomplishes both external and internal preference alignment: 1) It jointly integrates pairwise, pointwise, and contrastive preference alignment abilities into the reranker, achieving external preference alignment among RAG components. 2) It further introduces a pre-aligned stage before vanilla Supervised Fine-tuning (SFT), enabling LLMs to implicitly capture knowledge aligned with their reasoning preferences, achieving LLMs' internal alignment. Experimental results across four knowledge-intensive QA datasets demonstrate that DPA-RAG outperforms all baselines and seamlessly integrates both black-box and open-sourced LLM readers. Further qualitative analysis and discussions provide empirical guidance for achieving reliable RAG systems. Our code and example dataset are available at https://github.com/dongguanting/DPA-RAG.

Large language models (LLMs), with advanced linguistic capabilities, have been employed in reranking tasks through a sequence-to-sequence approach. In this paradigm, multiple passages are reranked in a listwise manner and a textual reranked permutation is generated. However, due to the limited context window of LLMs, this reranking paradigm requires a sliding window strategy to iteratively handle larger candidate sets. This not only increases computational costs but also restricts the LLM from fully capturing all the comparison information for all candidates. To address these challenges, we propose a novel self-calibrated listwise reranking method, which aims to leverage LLMs to produce global relevance scores for ranking. To achieve it, we first propose the relevance-aware listwise reranking framework, which incorporates explicit list-view relevance scores to improve reranking efficiency and enable global comparison across the entire candidate set. Second, to ensure the comparability of the computed scores, we propose self-calibrated training that uses point-view relevance assessments generated internally by the LLM itself to calibrate the list-view relevance assessments. Extensive experiments and comprehensive analysis on the BEIR benchmark and TREC Deep Learning Tracks demonstrate the effectiveness and efficiency of our proposed method.

Retrieval-Augmented Generation (RAG) has been shown to improve knowledge capabilities and alleviate the hallucination problem of LLMs. The Web is a major source of external knowledge used in RAG systems, and many commercial RAG systems have used Web search engines as their major retrieval systems. Typically, such RAG systems retrieve search results, download HTML sources of the results, and then extract plain texts from the HTML sources. Plain text documents or chunks are fed into the LLMs to augment the generation. However, much of the structural and semantic information inherent in HTML, such as headings and table structures, is lost during this plain-text-based RAG process. To alleviate this problem, we propose HtmlRAG, which uses HTML instead of plain text as the format of retrieved knowledge in RAG. We believe HTML is better than plain text in modeling knowledge in external documents, and most LLMs possess robust capacities to understand HTML. However, utilizing HTML presents new challenges. HTML contains additional content such as tags, JavaScript, and CSS specifications, which bring extra input tokens and noise to the RAG system. To address this issue, we propose HTML cleaning, compression, and a two-step block-tree-based pruning strategy, to shorten the HTML while minimizing the loss of information. Experiments on six QA datasets confirm the superiority of using HTML in RAG systems. Our code and datasets are available at https://github.com/plageon/HtmlRAG.

Multimodal learning is expected to boost model performance by integrating information from different modalities. However, its potential is not fully exploited because the widely-used joint training strategy, which has a uniform objective for all modalities, leads to imbalanced and under-optimized uni-modal representations. Specifically, we point out that there often exists modality with more discriminative information, e.g., vision of playing football and sound of blowing wind. They could dominate the joint training process, resulting in other modalities being significantly under-optimized. To alleviate this problem, we first analyze the under-optimized phenomenon from both the feed-forward and the back-propagation stages during optimization. Then, On-the-fly Prediction Modulation (OPM) and On-the-fly Gradient Modulation (OGM) strategies are proposed to modulate the optimization of each modality, by monitoring the discriminative discrepancy between modalities during training. Concretely, OPM weakens the influence of the dominant modality by dropping its feature with dynamical probability in the feed-forward stage, while OGM mitigates its gradient in the back-propagation stage. In experiments, our methods demonstrate considerable improvement across a variety of multimodal tasks. These simple yet effective strategies not only enhance performance in vanilla and task-oriented multimodal models, but also in more complex multimodal tasks, showcasing their effectiveness and flexibility.

Recently, Large Language Models (LLMs) such as ChatGPT have showcased remarkable abilities in solving general tasks, demonstrating the potential for applications in recommender systems. To assess how effectively LLMs can be used in recommendation tasks, our study primarily focuses on employing LLMs as recommender systems through prompt engineering. We propose a general framework for leveraging LLMs in recommendation tasks, focusing on the capabilities of LLMs as recommenders. To conduct our analysis, we formalize the input of LLMs for recommendation into natural language prompts with two key aspects and explain how our framework can be generalized to various recommendation scenarios. As for the use of LLMs as recommenders, we analyze the impact of public availability, tuning strategies, model architecture, parameter scale, and context length on recommendation results based on the classification of LLMs. As for prompt engineering, we further analyze the impact of four important components of prompts, i.e., task descriptions, user interest modeling, candidate items construction, and prompting strategies. In each section, we first define and categorize concepts in line with the existing literature. Then, we propose inspiring research questions followed by detailed experiments on two public datasets, in order to systematically analyze the impact of different factors on recommendation performance. Based on our empirical analysis, we finally summarize promising directions to shed lights on future research.

In search sessions, a series of interactions in the context has been proven to be advantageous in capturing users’ search intents. Existing studies show that designing pre-training tasks and data augmentation strategies for session search improves the robustness and generalizability of the model. However, such data augmentation strategies only focus on changing the original session structure to learn a better representation. Ignoring information from outside the session, users’ diverse and complex intents cannot be learned well by simply reordering and deleting historical behaviors, proving that such strategies are limited and inadequate. In order to solve the problem of insufficient modeling under complex user intents, we propose exploiting information outside the original session. More specifically, in this paper, we sample queries and documents from the global click-on and follow-up session graph, alter an original session with these samples, and construct a new session that shares a similar user intent with the original one. Specifically, we design four data augmentation strategies based on session graphs in view of both one-hop and multi-hop structures to sample intent-associated query/document nodes. Experiments conducted on three large-scale public datasets demonstrate that our model outperforms the existing ad-hoc and context-aware document ranking models.

Existing studies have proven that pre-trained ranking models outperform pre-trained language models when it comes to ranking tasks. To pre-train such models, researchers have utilized large-scale search logs and clicks as weak-supervised signals of query-document relevance. However, search logs are incomplete and sparse. Different users with the same intent tend to use various forms of queries. It is hard for recorded clicks to sufficiently cover diverse relevance patterns between queries and documents. Moreover, the diverse intentions of a large user base lead to long-tail distributions of search intents. Deriving sufficient relevance signals from sparse clicks of these long-tail intents poses another challenge. Therefore, there is significant potential for exploring richer relevance signals beyond direct clicks to pre-train high-quality ranking models. To tackle this problem, we develop two exploratory data augmentation strategies that consider the diversity of query forms from local and global perspectives, hence mining potential and diverse relevance signals from search logs. A generative augmentation strategy is also devised to create supplementary positive samples, to enhance the ranking ability for long-tail query intents. We leverage a multi-level pairwise ranking objective and a contrastive learning approach to enable our model to capture fine-grained relevance patterns and be robust for noisy training samples. Experimental results on a large-scale public dataset and a commercial dataset confirm that our model, namely PRADA, can yield better ranking effectiveness over existing pre-trained ranking models.

A key challenge in personalized product search is to capture user’s preferences. Recent work attempted to model sequences of user historical behaviors, i.e., product purchase histories, to build user profiles and to personalize results accordingly. Although these approaches have demonstrated promising retrieval performances, we notice that most of them focus solely on the intra-sequence interactions between items. However, as there is usually a small amount of historical behavior data, the user profiles learned by these approaches could be very sensitive to the noise included in it. To tackle this problem, we propose incorporating out-of-sequence external information to enhance user modeling. More specifically, we inject the external item–item relations (e.g., belonging to the same brand), and query–query relations (e.g., the semantic similarities between them), into the intra-sequence interaction to learn better user profiles. In addition, we devise two auxiliary decoders, with the historical item sequence reconstruction task and the global item similarity prediction task, to further improve the reliability of user modeling. Experimental results on two datasets from simulated and real user search logs respectively show that the proposed personalized product search method outperforms existing approaches.

Simulating high quality user behavior data has always been a fundamental yet challenging problem in human-centered applications such as recommendation systems, social networks, among many others. The major difficulty of user behavior simulation originates from the intricate mechanism of human cognitive and decision processes. Recently, substantial evidence has suggested that by learning huge amounts of web knowledge, large language models (LLMs) can achieve human-like intelligence and generalization capabilities. Inspired by such capabilities, in this article, we take an initial step to study the potential of using LLMs for user behavior simulation in the recommendation domain. To make LLMs act like humans, we design profile, memory and action modules to equip them, building LLM-based agents to simulate real users. To enable interactions between different agents and observe their behavior patterns, we design a sandbox environment, where each agent can interact with the recommendation system, and different agents can converse with their friends via one-to-one chatting or one-to-many social broadcasting. In the experiments, we first demonstrate the believability of the agent-generated behaviors based on both subjective and objective evaluations. Then, to show the potential applications of our method, we simulate and study two social phenomena including (1) information cocoons and (2) user conformity behaviors. We find that controlling the personalization degree of recommendation algorithms and improving the heterogeneity of user social relations can be two effective strategies for alleviating the problem of information cocoon, and the conformity behaviors can be highly influenced by the amount of user social relations. To advance this direction, we have released our project at https://github.com/RUC-GSAI/YuLan-Rec .

Multi-stakeholder recommender systems involve various roles, such as users and providers. Previous work pointed out that max-min fairness (MMF) is a better metric to support weak providers. However, when considering MMF, the features or parameters of these roles vary over time, and how to ensure long-term provider MMF has become a significant challenge. We observed that recommendation feedback loops (RFL) will influence the provider MMF greatly in the long term. RFL means that recommender systems can only receive feedback on exposed items from users and update recommender models incrementally based on this feedback. When utilizing the feedback, the recommender model will regard the unexposed items as negative. In this way, the tail provider will not get the opportunity to be exposed, and its items will always be considered negative samples. Such phenomena will become more and more serious in RFL. To alleviate the problem, this article proposes an online ranking model named Long-Term Provider Max-min Fairness (LTP-MMF). Theoretical analysis shows that the long-term regret of LTP-MMF enjoys a sub-linear bound. Experimental results on three public recommendation benchmarks demonstrated that LTP-MMF can outperform the baselines in the long term.

Large language model (LLM)-based agents have recently attracted much attention from the research and industry communities. Compared with original LLMs, LLM-based agents are featured in their self-evolving capability, which is the basis for solving real-world problems that need long-term and complex agent-environment interactions. The key component to support agent-environment interactions is the memory of the agents. While previous studies have proposed many promising memory mechanisms, they are scattered in different papers, and there lacks a systematical review to summarize and compare these works from a holistic perspective, failing to abstract common and effective designing patterns for inspiring future studies. To bridge this gap, in this article, we propose a comprehensive survey on the memory mechanism of LLM-based agents. In specific, we first discuss “what is” and “why do we need” the memory in LLM-based agents. Then, we systematically review previous studies on how to design and evaluate the memory module. In addition, we also present many agent applications, where the memory module plays an important role. At last, we analyze the limitations of existing work and show important future directions. To keep up with the latest advances in this field, we create a repository at https://github.com/nuster1128/LLM_Agent_Memory_Survey .

With the widespread deployment of recommender systems on various online platforms, researchers are striving to develop transferable recommendation algorithms that can effectively adapt to new task scenarios without requiring the re-training of new recommenders. However, there have been challenges in dealing with explicit ID modeling in this context. Recently, researchers have drawn inspiration from the achievements of pre-trained language models (PLMs), making it possible to acquire ID-agnostic representations by utilizing the corresponding texts of items. These representations have shown to be transferable across diverse domains. However, while these methods demonstrate generalization, they are less proficient in making personalized recommendations as they learn universal representation. In light of this issue, we present a review-enhanced universal sequence representation learning approach named RUNSRec . Our goal is to not only comprehend universal user behavioral patterns across different domains but also capture their inherent preferences to make recommendations. Our approach makes three technical advancements toward this objective. Firstly, we introduce a lightweight item encoding architecture based on parametric whitening and mixture-of-experts enhanced adapter. It learns discriminative item textual representations by encoding their corresponding identity text and review text, with a discriminative keyword extraction method to enhance the representation identifiability. Secondly, we propose a universal sequence representation learning method that enables the training of transferable recommenders across diverse domains, based on two novel contrastive learning tasks. Furthermore, we introduce a personalized adapter tuning mechanism that enables the universal recommender to capture user personal preferences in a parameter-efficient way. By incorporating universal behavioral patterns learned during the pre-training stage and personalized user tastes captured through adapter tuning, our approach achieves a better balance between generalization and personalization in transferable recommender systems. Extensive experiments conducted on five real-world datasets have demonstrated the effectiveness of our proposed approach.

In the past few decades, recommender systems have attracted much attention in both research and industry communities. Existing recommendation models mainly learn the underlying user preference from historical behavior data (typically in the forms of item IDs), and then estimate the user–item matching relationships for recommendations. Inspired by the recent progress on large language models (LLMs), we develop a different recommendation paradigm, considering recommendation as instruction following by LLMs. The key idea is that the needs of a user can be expressed in natural language descriptions (called instructions ), so that LLMs can understand and further execute the instruction for fulfilling the recommendation. For this purpose, we instruction tune the 3B Flan-T5-XL, to better adapt LLMs to recommender systems. We first design a general instruction format for describing the preference, intention, and task form of a user in natural language. Then we manually design 39 instruction templates and automatically generate large amounts of user-personalized instruction data with varying types of preferences and intentions. To demonstrate the effectiveness of our approach, we instantiate the instructions into several widely studied recommendation (or search) tasks, and conduct extensive experiments with real-world datasets. Experiment results show that our approach can outperform several competitive baselines, including the powerful GPT-3.5, on these evaluation tasks. Our approach sheds light on developing user-friendly recommender systems, in which users can freely communicate with the system and obtain accurate recommendations via natural language instructions.

We are again confronted with one of the most vexing aspects of the advancement of technology: automation and AI technology cause the devaluation of human labor, resulting in unemployment. With this background, automatic person-job fit systems are promising solutions to promote the employment rate. The purpose of person-job fit is to calculate a matching score between the job seeker's resume and the job posting, determining whether the job seeker is suitable for the position. In this paper, we propose a new approach to person-job fit that characterizes the hidden preference derived from the job seeker's career path. We categorize and utilize three types of preferences in the career path: consistency, likeness, and continuity. We prove that understanding the career path enables us to provide more appropriate career suggestions to job seekers. To demonstrate the practical value of our proposed model, we conduct extensive experiments on real-world data extracted from an online recruitment platform and then present detailed cases to show how the career path matter in person-job fit.

Yutao Zhu, Peitian Zhang, Chenghao Zhang, Yifei Chen, Binyu Xie, Zheng Liu, Ji-Rong Wen, Zhicheng Dou. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Changyu Chen, Xiting Wang, Ting-En Lin, Ang Lv, Yuchuan Wu, Xin Gao, Ji-Rong Wen, Rui Yan, Yongbin Li. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Junyi Li, Jie Chen, Ruiyang Ren, Xiaoxue Cheng, Xin Zhao, Jian-Yun Nie, Ji-Rong Wen. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Key-value (KV) caching is an important technique to accelerate the inference of large language models (LLMs), but incurs significant memory overhead.To compress the size of KV cache, existing methods often compromise precision or require extra data for calibration, limiting their practicality in LLM deployment.In this paper, we introduce DecoQuant, a novel data-free low-bit quantization technique based on tensor decomposition methods, to effectively compress KV cache.Our core idea is to adjust the outlier distribution of the original matrix by performing tensor decomposition, so that the quantization difficulties are migrated from the matrix to decomposed local tensors.Specially, we find that outliers mainly concentrate on small local tensors, while large tensors tend to have a narrower value range.Based on this finding, we propose to apply low-bit quantization to the large tensor, while maintaining high-precision representation for the small tensor.Furthermore, we utilize the proposed quantization method to compress the KV cache of LLMs to accelerate the inference and develop an efficient dequantization kernel tailored specifically for DecoQuant.Through extensive experiments, DecoQuant demonstrates remarkable efficiency gains, showcasing up to a ∼75% reduction in memory footprint while maintaining comparable generation quality.

Hongda Sun, Weikai Xu, Wei Liu, Jian Luan, Bin Wang, Shuo Shang, Ji-Rong Wen, Rui Yan. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

The integration of large language models (LLMs) and search engines represents a significant evolution in knowledge acquisition methodologies.However, determining the knowledge that an LLM already possesses and the knowledge that requires the help of a search engine remains an unresolved issue.Most existing methods solve this problem through the results of preliminary answers or reasoning done by the LLM itself, but this incurs excessively high computational costs.This paper introduces a novel collaborative approach, namely SlimPLM, that detects missing knowledge in LLMs with a slim proxy model, to enhance the LLM's knowledge acquisition process.We employ a proxy model which has far fewer parameters, and take its answers as heuristic answers.Heuristic answers are then utilized to predict the knowledge required to answer the user question, as well as the known and unknown knowledge within the LLM.We only conduct retrieval for the missing knowledge in questions that the LLM does not know.Extensive experimental results on five datasets with two LLMs demonstrate a notable improvement in the end-to-end performance of LLMs in question-answering tasks, achieving or surpassing current state-of-the-art models with lower LLM inference costs. 1

Tianyi Tang, Wenyang Luo, Haoyang Huang, Dongdong Zhang, Xiaolei Wang, Xin Zhao, Furu Wei, Ji-Rong Wen. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Recently, integrating external tools with Large Language Models (LLMs) has gained significant attention as an effective strategy to mitigate the limitations inherent in their pre-training data. However, real-world systems often incorporate a wide array of tools, making it impractical to input all tools into LLMs due to length limitations and latency constraints. Therefore, to fully exploit the potential of tool-augmented LLMs, it is crucial to develop an effective tool retrieval system. Existing tool retrieval methods primarily focus on semantic matching between user queries and tool descriptions, frequently leading to the retrieval of redundant, similar tools. Consequently, these methods fail to provide a complete set of diverse tools necessary for addressing the multifaceted problems encountered by LLMs. In this paper, we propose a novel modelagnostic CO llaborative L earning-based T ool Retrieval approach, COLT, which captures not only the semantic similarities between user queries and tool descriptions but also takes into account the collaborative information of tools. Specifically, we first fine-tune the PLM-based retrieval models to capture the semantic relationships between queries and tools in the semantic learning stage. Subsequently, we construct three bipartite graphs among queries, scenes, and tools and introduce a dual-view graph collaborative learning framework to capture the intricate collaborative relationships among tools during the collaborative learning stage. Extensive experiments on both the open benchmark and the newly introduced ToolLens dataset show that COLT achieves superior performance. Notably, the performance of BERT-mini (11M) with our proposed model framework outperforms BERT-large (340M), which has 30 times more parameters. Furthermore, we will release ToolLens publicly to facilitate future research on tool retrieval.

Chain-of-Thought (CoT) prompting can enhance the reasoning capabilities of large language models (LLMs), establishing itself as a primary approach to solving complex reasoning tasks. Existing CoT synthesis approaches usually focus on simpler reasoning tasks and thus result in low-quality and inconsistent CoT prompts. In response to this challenge, we present an empirical investigation of CoT prompting and introduce CoTGenius, a novel framework designed for the automatic generation of superior CoT prompts. CoTGenius is developed based on three major evolution strategies, i.e., complicate, diversify, and specify—alongside two filtering mechanisms: evolutionary success judgement and correctness verification. We further employ CoTGenius to create an extensive CoT dataset, and subsequently fine-tune the Llama 2-Chat 7B and 13B models on this dataset. We call the resulting model ChainLM. To deal with the cumulative error issue in reasoning steps, we propose a step-level debating method, wherein multiple debaters discuss each reasoning step to arrive at the correct answer. Extensive experiments demonstrate that our ChainLM models exhibit enhanced proficiency in addressing a spectrum of complex reasoning problems compared to existing models. In addition, we conduct an in-depth analysis of the impact of data categories within CoTGenius on the model performance. We release our dataset and code at https://github.com/RUCAIBox/ChainLM.

Large language models (LLMs) have achieved dramatic proficiency over NLP tasks with normal length. Recently, multiple studies have committed to extending the context length and enhancing the long text modeling capabilities of LLMs. To comprehensively evaluate the long context ability of LLMs, we propose BAMBOO, a multi-task long context benchmark. BAMBOO has been designed with four principles: comprehensive capacity evaluation, avoidance of data contamination, accurate automatic evaluation, and different length levels. It consists of 10 datasets from 5 different long text understanding tasks, i.e., question answering, hallucination detection, text sorting, language modeling, and code completion, to cover various domains and core capacities of LLMs. We conduct experiments with five widely-used long-context models and further discuss five key questions for long text research. In the end, we discuss problems of current long-context models and point out future directions for enhancing long text modeling capacities. We release our data, prompts, and code at https://anonymous.4open.science/r/BAMBOO/.

Lightweight fine-tuning is widely used as an important technique for efficiently adapting pre-trained language models (PLM) to downstream tasks. Despite the reduction in trainable parameters, existing lightweight fine-tuning methods are found to be effective in low-resource settings but often fail in high-resource settings, leading to unreliable outcomes. This limitation can be attributed to inflexible strategies: they identify the parameters of the model to be trained before fine-tuning and remain unchanged without taking into account the inherent variance of generalization ability in model components (i.e., feed-forward, attention layers) and potential changes during the fine-tuning process. In this paper, we introduce a simple but effective calibration for lightweight fine-tuning PLMs based on the matrix’s stable rank according to both model components and the training process. We proposed both theoretical analyses and experimental verification for the proposed calibration strategy. Considering efficiency, we further propose time-aware and structure-aware strategies to determine the most crucial time to commence the fine-tuning procedure and selectively apply parameter matrices for lightweight fine-tuning, respectively. Extensive experiments demonstrate the superiority of our proposed fine-tuning approach (average improvement 3.1 for GLUE score compared to lightweight fine-tuning method).

Despite the superior performance, Large Language Models (LLMs) require significant computational resources for deployment and use. To overcome this issue, quantization methods have been widely applied to reduce the memory footprint of LLMs as well as increase the inference rate. However, a major challenge is that low-bit quantization methods often lead to performance degradation. It is important to understand how quantization impacts the capacity of LLMs. Different from previous studies focused on overall performance, this work aims to investigate the impact of quantization on emergent abilities, which are important characteristics that distinguish LLMs from small language models. Specifically, we examine the abilities of in-context learning, chain-of-thought reasoning, and instruction-following in quantized LLMs. Our empirical experiments show that these emergent abilities still exist in 4-bit quantization models, while 2-bit models encounter severe performance degradation on the test of these abilities. To improve the performance of low-bit models, we conduct two special experiments: (1) fine-gained impact analysis that studies which components (or substructures) are more sensitive to quantization, and (2) performance compensation through model fine-tuning. Our work derives a series of important findings to understand the impact of quantization on emergent abilities and sheds light on the possibilities of extremely low-bit quantization for LLMs.

Text-to-image generative models, specifically those based on diffusion models like Imagen and Stable Diffusion, have made substantial advancements. Recently, there has been a surge of interest in the delicate refinement of text prompts. Users assign weights or alter the injection time steps of certain words in the text prompts to improve the quality of generated images. However, the success of fine-control prompts depends on the accuracy of the text prompts and the careful selection of weights and time steps, which requires significant manual intervention. To address this, we introduce the Prompt Auto-Editing (PAE) method. Besides refining the original prompts for image generation, we further employ an online reinforcement learning strategy to explore the weights and injection time steps of each word, leading to the dynamic fine-control prompts. The re-wardfunction during training encourages the model to consider aesthetic score, semantic consistency, and user prefer-ences. Experimental results demonstrate that our proposed method effectively improves the original prompts, generating visually more appealing images while maintaining semantic alignment. Code is available at this https URL.

Recently, continuous diffusion models (CDM) have been introduced into non-autoregressive (NAR) text-to-text generation.However, the discrete nature of text increases the difficulty of CDM to generate coherent and fluent texts, and also causes the incompatibility problem between CDM and advanced NLP techniques, especially the popular pre-trained language models (PLMs).To solve it, we propose Diffusion-NAT, which introduces discrete diffusion models (DDM) into NAR text-to-text generation and integrates BART to improve the performance.By revising the decoding process of BART and the typical settings of DDM, we unify the inference process of BART and the denoising process of DDM into the same NAR masked tokens recovering task.In this way, DDM can rely on BART to perform denoising, which can benefit from both the rich prelearned knowledge of BART and the iterative refining paradigm of DDM.Besides, we also propose the iterative self-prompting strategy to further improve the generation quality.Experimental results on 7 datasets show that our approach can outperform competitive NAR methods, and even surpass autoregressive methods.Our code and data are released at

Large language models (LLMs) are still struggling in aligning with human preference in complex tasks and scenarios.They are prone to overfit into the unexpected patterns or superficial styles in the training data.We conduct an empirical study that only selects the top-10% most updated parameters in LLMs for alignment training, and see improvements in the convergence process and final performance.It indicates the existence of redundant neurons in LLMs for alignment training.To reduce its influence, we propose a low-redundant alignment method named ALLO, focusing on optimizing the most related neurons with the most useful supervised signals.Concretely, we first identify the neurons that are related to the human preference data by a gradient-based strategy, then identify the alignment-related key tokens by reward models for computing loss.Besides, we also decompose the alignment process into the forgetting and learning stages, where we first forget the tokens with unaligned knowledge and then learn aligned knowledge, by updating different ratios of neurons, respectively.Experimental results on 10 datasets have shown the effectiveness of ALLO.Our code and data are available at https://github.com/RUCAIBox/ALLO.

Hallucination detection is a challenging task for large language models (LLMs), and existing studies heavily rely on powerful closed-source LLMs such as GPT-4.In this paper, we propose an autonomous LLM-based agent framework, called HaluAgent, which enables relatively smaller LLMs 1 (e.g., Baichuan2-Chat 7B) to actively select suitable tools for detecting multiple hallucination types such as text, code, and mathematical expression.In HaluAgent, we integrate the LLM, multi-functional toolbox, and design a fine-grained three-stage detection framework along with memory mechanism.To facilitate the effectiveness of HaluAgent, we leverage existing Chinese and English datasets to synthesize detection trajectories for fine-tuning, which endows HaluAgent with the capability for bilingual hallucination detection.Extensive experiments demonstrate that only using 2K samples for tuning LLMs, HaluAgent can perform hallucination detection on various types of tasks and datasets, achieving performance comparable to or even higher than GPT-4 without tool enhancements on both indomain and out-of-domain datasets.We release our dataset and code at https://github.com/ RUCAIBox/HaluAgent.

Recent research observed a noteworthy phenomenon in large language models (LLMs), referred to as the "reversal curse."The reversal curse is that when dealing with two entities, denoted as a and b, connected by their relation R and its inverse R -1 , LLMs excel in handling sequences in the form of "aRb," but encounter challenges when processing "bR -1 a," whether in generation or comprehension.For instance, GPT-4 can accurately respond to the query "Tom Cruise's mother is?" with "Mary Lee Pfeiffer," but it struggles to provide a satisfactory answer when asked "Mary Lee Pfeiffer's son is?"In this paper, we undertake the first-ever study of how the reversal curse happens in LLMs.Our investigations reveal that the reversal curse can stem from the specific training objectives, which become particularly evident in the widespread use of next-token prediction within most causal language models.We hope this initial investigation can draw more attention to the reversal curse, as well as other underlying limitations in current LLMs. 1

Considering the limited internal parametric knowledge, retrieval-augmented generation (RAG) has been widely used to extend the knowledge scope of large language models (LLMs).Despite the extensive efforts on RAG research, in existing methods, LLMs cannot precisely assess the relevance of retrieved documents, thus likely leading to misleading or even incorrect utilization of external knowledge (i.e., retrieved documents).To address this issue, in this paper, we propose REAR, a RElevance-Aware Retrievalaugmented approach for open-domain question answering (QA).As the key motivation, we aim to enhance the self-awareness regarding the reliability of external knowledge for LLMs, so as to adaptively utilize external knowledge in RAG systems.Specially, we develop a novel architecture for LLM-based RAG systems, by incorporating a specially designed assessment module that precisely assesses the relevance of retrieved documents.Furthermore, we propose an improved training method based on bigranularity relevance fusion and noise-resistant training.By combining the improvements in both architecture and training, our proposed REAR can better utilize external knowledge by effectively perceiving the relevance of retrieved documents.Experiments on four opendomain QA tasks show that REAR significantly outperforms previous a number of competitive RAG approaches.

With the popularity of online music services, personalized music recommendation has garnered much research interest. Recommendation models are typically trained on datasets constructed from user feedback, which includes both the active feedback (e.g., clicking the Like or Skip buttons) and passive feedback (e.g., auto-play), with passive feedback comprising the majority. Due to the unavailability of user attention, the massive amount of passive feedback is unreliable, significantly compromising the quality of the training data. How to estimate the user's attention on the target music has become a critical problem in music recommendation. Heuristic methods such as exponential decay and negative sampling have been proposed. However, they either neglect the sequential dependencies between feedback actions or utilize only a small fraction of passive samples, leading to inaccurate and biased attention estimation. In this paper, we naturally propose modeling user attention prediction as a positive-unlabeled (PU) learning problem, where active feedback is treated as positive samples and passive feedback is treated as unlabeled samples, as we can only ensure that the user's attention is focused when she provides active feedback. Then we propose an extended PU-learning model with sequential dependencies, called UAE, which contains an unbiased user attention estimator and an unbiased propensity estimator. Subsequently, a joint learning algorithm is developed in which the attention and propensity estimators are optimized in alternating fashion. Theoretical analysis shows the unbiasedness and variance of the attention estimator and the propensity estimator. Extensive experiments on two large-scale datasets demonstrate the proposed UAE's effectiveness and generality in enhancing downstream music recommendation. One week online A/B testing on Huawei Music App manifests that UAE can significantly increase the users' play count and time over 2%, further demonstrating the effectiveness of UAE in real-world music recommendation products.

Recently, geometric-inspired embedding methods draw research interests for their superior ability in representing transitive and asymmetric relations. A typical example, box embeddings, in which objects are parameterized as axis-aligned hyper-rectangles (i.e. boxes), can effectively model the partial orders and similarities between objects with the inclusion and overlapping relations of the boxes. However, the hard edges of the boxes present difficulties for gradient-based optimization. In this paper, we first identify two problems that may hinder the optimization of box embeddings, namely the zero-gradient problem and the vanishing gradient problem in high dimensional settings. Then, we propose a simple yet effective framework (Box++) to tackle these problems. For the first problem, Box++ combines the overlapping volumes and distance measures in optimizing box embeddings. The distance measures can naturally measure the “degree of disjointedness” for disjoint boxes and provide reasonable gradients for optimization. For the second problem, we theoretically prove that under certain conditions, the gradient would vanish exponentially, and therefore, make the optimization converges to suboptimal solutions. We further design a gradient normalization strategy for the Box++ models to alleviate this problem by rescaling the vanishing gradient. Extensive experiments on real-world datasets demonstrate the effectiveness of the proposed framework.

Recently, large language models (LLMs) have shown great potential in recommender systems, either improving existing recommendation models or serving as the backbone. However, there exists a large semantic gap between LLMs and recommender systems, since items to be recommended are often indexed by discrete identifiers (item ID) out of the LLM's vocabulary. In essence, LLMs capture language semantics while recommender systems imply collaborative semantics, making it difficult to sufficiently leverage the model capacity of LLMs for recommendation. To address this challenge, in this paper, we propose a new LLM-based recommendation model called LC-Rec, which can better integrate language and collaborative semantics for recommender systems. Our approach can directly generate items from the entire item set for recommendation, without relying on candidate items. Specifically, we make two major contributions in our approach. For item indexing, we design a learning-based vector quantization method with uniform semantic mapping, which can assign meaningful and non-conflicting IDs (called item indices) for items. For alignment tuning, we propose a series of specially designed tuning tasks to enhance the integration of collaborative semantics in LLMs. Our fine-tuning tasks enforce LLMs to deeply integrate language and collaborative semantics (characterized by the learned item indices), so as to achieve an effective adaptation to recommender systems. Extensive experiments demonstrate the effectiveness of our method, showing that our approach can outperform a number of competitive baselines including traditional recommenders and existing LLM-based recommenders. Our code is available at https://github.com/RUCAIBox/LC-Rec/.

Personalized PageRank (PPR) is an extensively studied and applied node proximity measure in graphs. For a pair of nodes $s$ and $t$ on a graph $G=(V,E)$, the PPR value $π(s,t)$ is defined as the probability that an $α$-discounted random walk from $s$ terminates at $t$, where the walk terminates with probability $α$ at each step. We study the classic Single-Source PPR query, which asks for PPR approximations from a given source node $s$ to all nodes in the graph. Specifically, we aim to provide approximations with absolute error guarantees, ensuring that the resultant PPR estimates $\hatπ(s,t)$ satisfy $\max_{t\in V}\big|\hatπ(s,t)-π(s,t)\big|\le\varepsilon$ for a given error bound $\varepsilon$. We propose an algorithm that achieves this with high probability, with an expected running time of - $\widetilde{O}\big(\sqrt{m}/\varepsilon\big)$ for directed graphs, where $m=|E|$; - $\widetilde{O}\big(\sqrt{d_{\mathrm{max}}}/\varepsilon\big)$ for undirected graphs, where $d_{\mathrm{max}}$ is the maximum node degree in the graph; - $\widetilde{O}\left(n^{γ-1/2}/\varepsilon\right)$ for power-law graphs, where $n=|V|$ and $γ\in\left(\frac{1}{2},1\right)$ is the extent of the power law. These sublinear bounds improve upon existing results. We also study the case when degree-normalized absolute error guarantees are desired, requiring $\max_{t\in V}\big|\hatπ(s,t)/d(t)-π(s,t)/d(t)\big|\le\varepsilon_d$ for a given error bound $\varepsilon_d$, where the graph is undirected and $d(t)$ is the degree of node $t$. We give an algorithm that provides this error guarantee with high probability, achieving an expected complexity of $\widetilde{O}\left(\sqrt{\sum_{t\in V}π(s,t)/d(t)}\big/\varepsilon_d\right)$. This improves over the previously known $O(1/\varepsilon_d)$ complexity.

Alignment with human preference is a desired property of large language models (LLMs). Currently, the main alignment approach is based on reinforcement learning from human feedback (RLHF). Despite the effectiveness of RLHF, it is intricate to implement and train, thus recent studies explore how to develop alternative alignment approaches based on supervised fine-tuning (SFT). A major limitation of SFT is that it essentially does imitation learning, which can't fully understand what are the expected behaviors. To address this issue, we propose an improved alignment approach named $\textbf{FIGA}$. Different from prior methods, we incorporate fine-grained (i.e., token or phrase level) quality signals that are derived by contrasting good and bad responses. Our approach has made two major contributions. Firstly, we curate a refined alignment dataset that pairs initial responses and the corresponding revised ones. Secondly, we devise a new loss function can leverage fine-grained quailty signals to instruct the learning of LLMs for alignment. Extensive experiments have demonstrated the effectiveness of our approaches by comparing a number of competitive baselines.

Medical vision-language pre-training methods mainly leverage the correspondence between paired medical images and radiological reports. Although multi-view spatial images and temporal sequences of image-report pairs are available in off-the-shelf multi-modal medical datasets, most existing methods have not thoroughly tapped into such extensive supervision signals. In this paper, we introduce the Med-ST framework for fine-grained spatial and temporal modeling to exploit information from multiple spatial views of chest radiographs and temporal historical records. For spatial modeling, Med-ST employs the *Mixture of View Expert (MoVE)* architecture to integrate different visual features from both frontal and lateral views. To achieve a more comprehensive alignment, Med-ST not only establishes the global alignment between whole images and texts but also introduces modality-weighted local alignment between text tokens and spatial regions of images. For temporal modeling, we propose a novel cross-modal bidirectional cycle consistency objective by forward mapping classification (FMC) and reverse mapping regression (RMR). By perceiving temporal information from simple to complex, Med-ST can learn temporal semantics. Experimental results across four distinct tasks demonstrate the effectiveness of Med-ST, especially in temporal classification tasks. Our code and model are available at https://github.com/SVT-Yang/MedST.

Feature interaction learning (FIL) focuses on capturing the complex relationships among multiple features for building predictive models, which is widely used in real-world tasks. Despite the research progress, existing FIL methods suffer from two major limitations. Firstly, they mainly model the feature interactions within a bounded order (e.g., small integer order) due to the exponential growth of the interaction terms. Secondly, the interaction order of each feature is often independently learned, which lacks the flexibility to capture the feature dependencies in varying contexts.

In video recommendation, an ongoing effort is to satisfy users' personalized information needs by leveraging their logged watch time. However, watch time prediction suffers from duration bias, hindering its ability to reflect users' interests accurately. Existing label-correction approaches attempt to uncover user interests through grouping and normalizing observed watch time according to video duration. Although effective to some extent, we found that these approaches regard completely played records (i.e., a user watches the entire video) as equally high interest, which deviates from what we observed on real datasets: users have varied explicit feedback proportion when completely playing videos. In this paper, we introduce the counterfactual watch time (CWT), the potential watch time a user would spend on the video if its duration is sufficiently long. Analysis shows that the duration bias is caused by the truncation of CWT due to the video duration limitation, which usually occurs on those completely played records. Besides, a Counterfactual Watch Model (CWM) is proposed, revealing that CWT equals the time users get the maximum benefit from video recommender systems. Moreover, a cost-based transform function is defined to transform the CWT into the estimation of user interest, and the model can be learned by optimizing a counterfactual likelihood function defined over observed user watch times. Extensive experiments on three real video recommendation datasets and online A/B testing demonstrated that CWM effectively enhanced video recommendation accuracy and counteracted the duration bias.

Mathematical reasoning is an important capability of large language models~(LLMs) for real-world applications. To enhance this capability, existing work either collects large-scale math-related texts for pre-training, or relies on stronger LLMs (\eg GPT-4) to synthesize massive math problems. Both types of work generally lead to large costs in training or synthesis. To reduce the cost, based on open-source available texts, we propose an efficient way that trains a small LLM for math problem synthesis, to efficiently generate sufficient high-quality pre-training data. To achieve it, we create a dataset using GPT-4 to distill its data synthesis capability into the small LLM. Concretely, we craft a set of prompts based on human education stages to guide GPT-4, to synthesize problems covering diverse math knowledge and difficulty levels. Besides, we adopt the gradient-based influence estimation method to select the most valuable math-related texts. The both are fed into GPT-4 for creating the knowledge distillation dataset to train the small LLM. We leverage it to synthesize 6 million math problems for pre-training our JiuZhang3.0 model. The whole process only needs to invoke GPT-4 API 9.3k times and use 4.6B data for training. Experimental results have shown that JiuZhang3.0 achieves state-of-the-art performance on several mathematical reasoning datasets, under both natural language reasoning and tool manipulation settings. Our code and data will be publicly released in \url{https://github.com/RUCAIBox/JiuZhang3.0}.

Benefiting from the powerful language expression and planning capabilities of Large Language Models (LLMs), LLM-based autonomous agents have achieved promising performance in various downstream tasks. Recently, based on the development of single-agent systems, researchers propose to construct LLM-based multi-agent systems to tackle more complicated tasks. In this paper, we propose a novel framework, named COPPER, to enhance the collaborative capabilities of LLM-based agents with the self-reflection mechanism. To improve the quality of reflections, we propose to fine-tune a shared reflector, which automatically tunes the prompts of actor models using our counterfactual PPO mechanism. On the one hand, we propose counterfactual rewards to assess the contribution of a single agent’s reflection within the system, alleviating the credit assignment problem. On the other hand, we propose to train a shared reflector, which enables the reflector to generate personalized reflections according to agent roles, while reducing the computational resource requirements and improving training stability. We conduct experiments on three datasets to evaluate the performance of our model in multi-hop question answering, mathematics, and chess scenarios. Experimental results show that COPPER possesses stronger reflection capabilities and exhibits excellent generalization performance across different actor models.

Transformer-based large language models (LLMs) typically have a limited context window, resulting in significant performance degradation when processing text beyond the length of the context window. Extensive studies have been proposed to extend the context window and achieve length extrapolation of LLMs, but there is still a lack of in-depth interpretation of these approaches. In this study, we explore the positional information within and beyond the context window for deciphering the underlying mechanism of LLMs. By using a mean-based decomposition method, we disentangle positional vectors from hidden states of LLMs and analyze their formation and effect on attention. Furthermore, when texts exceed the context window, we analyze the change of positional vectors in two settings, i.e., direct extrapolation and context window extension. Based on our findings, we design two training-free context window extension methods, positional vector replacement and attention window extension. Experimental results show that our methods can effectively extend the context window length.

Standard Large Language Models (LLMs) struggle with handling dialogues with long contexts due to efficiency and consistency issues. According to our observation, dialogue contexts are highly structured, and the special token of End-of-Utterance (EoU) in dialogues has the potential to aggregate information. We refer to the EoU tokens as ``conversational attention sinks'' (conv-attn sinks). Accordingly, we introduce StreamingDialogue, which compresses long dialogue history into conv-attn sinks with minimal losses, and thus reduces computational complexity quadratically with the number of sinks (i.e., the number of utterances). Current LLMs already demonstrate the ability to handle long context window, e.g., a window size of 200K or more. To this end, by compressing utterances into EoUs, our method has the potential to handle more than 200K of utterances, resulting in a prolonged dialogue learning. In order to minimize information losses from reconstruction after compression, we design two learning strategies of short-memory reconstruction (SMR) and long-memory reactivation (LMR). Our method outperforms strong baselines in dialogue tasks and achieves a 4 $\times$ speedup while reducing memory usage by 18 $\times$ compared to dense attention recomputation.

When solving partial differential equations (PDEs), classical numerical methods often require fine mesh grids and small time stepping to meet stability, consistency, and convergence conditions, leading to high computational cost. Recently, machine learning has been increasingly utilized to solve PDE problems, but they often encounter challenges related to interpretability, generalizability, and strong dependency on rich labeled data. Hence, we introduce a new PDE-Preserved Coarse Correction Network (P$^2$C$^2$Net) to efficiently solve spatiotemporal PDE problems on coarse mesh grids in small data regimes. The model consists of two synergistic modules: (1) a trainable PDE block that learns to update the coarse solution (i.e., the system state), based on a high-order numerical scheme with boundary condition encoding, and (2) a neural network block that consistently corrects the solution on the fly. In particular, we propose a learnable symmetric Conv filter, with weights shared over the entire model, to accurately estimate the spatial derivatives of PDE based on the neural-corrected system state. The resulting physics-encoded model is capable of handling limited training data (e.g., 3--5 trajectories) and accelerates the prediction of PDE solutions on coarse spatiotemporal grids while maintaining a high accuracy. P$^2$C$^2$Net achieves consistent state-of-the-art performance with over 50\% gain (e.g., in terms of relative prediction error) across four datasets covering complex reaction-diffusion processes and turbulent flows.

At present, most recommender systems involve two stakeholders, providers and customers. Apart from maximizing the recommendation accuracy, the fairness issue for both sides should also be considered. Most of previous studies try to improve two-sided fairness with post-processing algorithms or fairness-aware loss constraints, which are highly dependent on the heuristic adjustments without respect to the optimization goal of accuracy. In contrast, we propose a novel training framework, adaptive weighting towards two-sided fairness-aware recommendation (named Ada2Fair), which lies in the extension of the accuracy-focused objective to a controllable preference learning loss over the interaction data. Specifically, we adjust the optimization scale of an interaction sample with an adaptive weight generator, and estimate the two-sided fairness-aware weights within model training. During the training process, the recommender is trained with two-sided fairness-aware weights to boost the utility of niche providers and inactive customers in a unified way. Extensive experiments on three public datasets verify the effectiveness of Ada2Fair, which can achieve Pareto efficiency in two-sided fairness-aware recommendation.

Scaling of neural networks has recently shown great potential to improve the model capacity in various fields. Specifically, model performance has a power-law relationship with model size or data size, which provides important guidance for the development of large-scale models. However, there is still limited understanding on the scaling effect of user behavior models in recommender systems, where the unique data characteristics (e.g., data scarcity and sparsity) pose new challenges in recommendation tasks.

The vital goal of information retrieval today extends beyond merely connecting users with relevant information they search for. It also aims to enrich the diversity, personalization, and interactivity of that connection, ensuring the information retrieval process is as seamless, beneficial, and supportive as possible in the global digital era. Current information retrieval systems often encounter challenges like a constrained understanding of queries, static and inflexible responses, limited personalization, and restricted interactivity. With the advent of large language models (LLMs), there's a transformative paradigm shift as we integrate LLM-powered agents into these systems. These agents bring forth crucial human capabilities like memory and planning to make them behave like humans in completing various tasks, effectively enhancing user engagement and offering tailored interactions. In this tutorial, we delve into the cutting-edge techniques of LLM-powered agents across various information retrieval fields, such as search engines, social networks, recommender systems, and conversational assistants. We will also explore the prevailing challenges in seamlessly incorporating these agents and hint at prospective research avenues that can revolutionize the way of information retrieval.

To capture user preference, transformer models have been widely applied to model sequential user behavior data. The core of transformer architecture lies in the self-attention mechanism, which computes the pairwise attention scores in a sequence. Due to the permutation-equivariant nature, positional encoding is used to enhance the attention between token representations. In this setting, the pairwise attention scores can be derived by both semantic difference and positional difference. However, prior studies often model the two kinds of difference measurements in different ways, which potentially limits the expressive capacity of sequence modeling.

We revisit ApproxContributions, the classic local graph exploration algorithm proposed by Andersen, Borgs, Chayes, Hopcroft, Mirrokni, and Teng (WAW ’07, Internet Math. ’08) for computing an є-approximation of the PageRank contribution vector for a target node t on a graph with n nodes and m edges. We give a worst-case complexity bound of it as O(nπ(t)/є·min(Δin,Δout,√m)), where π(t) is the PageRank score of t, and Δin and Δout are the maximum in-degree and out-degree of the graph, resp. We also give a lower bound of Ω(min(Δin/δ,Δout/δ,√m/δ,m)) for detecting t’s δ-contributing set, showing that the simple ApproxContributions algorithm is already optimal.

Search result diversification aims to provide a diversified document ranking list so as to cover as many intents as possible and satisfy the various information needs of different users. Existing approaches usually represented documents by pretrained embeddings (such as doc2vec and Glove). These document representations cannot adequately represent the document's content and are hard to capture the intrinsic user's intent coverage of the given query. Moreover, the limited number of labeled data for search result diversification exacerbates the difficulty of obtaining more efficient document representations. To alleviate these problems and learn more effective document representations, we propose a Contrastive Learning framework for search result DIVersification (CL4DIV). Specifically, we design three contrastive learning tasks from the perspective of subtopics, documents, and candidate document sequences, which correspond to three essential elements in search result diversification. These training tasks are employed to pretrain the document encoder and the document sequence encoder, which are used in the diversified ranking model. Experimental results show that øurs significantly outperforms all existing diversification models. Further analysis demonstrates that our method has wide applicability and can also be used to improve several existing methods.

Abstract Autonomous agents have long been a research focus in academic and industry communities. Previous research often focuses on training agents with limited knowledge within isolated environments, which diverges significantly from human learning processes, and makes the agents hard to achieve human-like decisions. Recently, through the acquisition of vast amounts of Web knowledge, large language models (LLMs) have shown potential in human-level intelligence, leading to a surge in research on LLM-based autonomous agents. In this paper, we present a comprehensive survey of these studies, delivering a systematic review of LLM-based autonomous agents from a holistic perspective. We first discuss the construction of LLM-based autonomous agents, proposing a unified framework that encompasses much of previous work. Then, we present a overview of the diverse applications of LLM-based autonomous agents in social science, natural science, and engineering. Finally, we delve into the evaluation strategies commonly used for LLM-based autonomous agents. Based on the previous studies, we also present several challenges and future directions in this field.

Modeling contextual information in a search session has drawn more and more attention when understanding complex user intents. Recent methods are all data-driven, i.e., they train different models on large-scale search log data to identify the relevance between search contexts and candidate documents. The common training paradigm is to pair the search context with different candidate documents and train the model to rank the clicked documents higher than the unclicked ones. However, this paradigm neglects the symmetric nature of the relevance between the session context and document, i.e., the clicked documents can also be paired with different search contexts when training. In this work, we propose query-oriented data augmentation to enrich search logs and empower the modeling. We generate supplemental training pairs by altering the most important part of a search context, i.e., the current query, and train our model to rank the generated sequence along with the original sequence. This approach enables models to learn that the relevance of a document may vary as the session context changes, leading to a better understanding of users’ search patterns. We develop several strategies to alter the current query, resulting in new training data with varying degrees of difficulty. Through experimentation on two extensive public search logs, we have successfully demonstrated the effectiveness of our model.

Personalized search and search result diversification are two possible solutions to cope with the query ambiguity problem in search engines. In most existing studies, they have been investigated separately, but intuitively, they address the problem from two complementary perspectives and should be combined. Some recent work tried to combine them by restricting result diversification to the subtopics corresponding to the user's personal profile. However, diversification can be required even when the subtopics are outside the user's profile. In this paper, we propose a more general approach to integrate them based on users' implicit feedback in query logs. The proposed approach PER+DIV aggregates a document's novelty score and personal relevance score dynamically according to how much the query falls into the user's interests. To train the model based on user clicks in the logs, we consider user click as a result of both personal relevance and result diversity and a new method is proposed to isolate and model these two factors. To evaluate the model, we design several diversified and personalized metrics in addition to the traditional click-based metrics. Experimental results on a large-scale query log dataset show that the proposed integrated method significantly outperforms the existing personalization and diversification approaches.

Providing human-understandable explanations for the matching predictions is still challenging for current legal case matching methods. One difficulty is that legal cases are semi-structured text documents with complicated case-case and case-law article correlations. To tackle the issue, we propose a novel graph optimal transport (GOT)-based legal case matching model that is able to provide not only the matching predictions but also plausible and faithful explanations for the prediction. The model, called GEIOT-Match, first constructs a heterogeneous graph to explicitly represent the semi-structured nature of legal cases and their associations with the law articles. Therefore, matching two legal cases amounts to identifying the rationales from the paired legal case sub-graphs in the heterogeneous graph and then aligning between them. An inverse optimal transport (IOT) model on graphs is learned to extract rationales from paired legal cases. The extracted rationales and the heterogeneous graph demonstrate the key legal characteristics of legal cases, which can be further used to conduct matching and generate explanations for the matching. Experimental results showed that GEIOT-Match outperformed state-of-the-art baselines in terms of matching prediction, rationale extraction, and natural language explanation generation.

Personalized PageRank (PPR) is a traditional measure for node proximity on large graphs. For a pair of nodes$\boldsymbol{s}$and$\boldsymbol{t}$, the PPR value$\boldsymbol{\pi_{s}(t)}$equals the probability that an$\boldsymbol{\alpha }$-discounted random walk from$\boldsymbol{s}$terminates at$\boldsymbol{t}$and reflects the importance between$\boldsymbol{s}$and$\boldsymbol{t}$in a bidirectional way. As a generalization of Google's celebrated PageRank centrality, PPR has been extensively studied and has found multifaceted applications in many fields, such as network analysis, graph mining, and graph machine learning. Despite numerous studies devoted to PPR over the decades, efficient computation of PPR remains a challenging problem, and there is a dearth of systematic summaries and comparisons of existing algorithms. In this paper, we recap several frequently used techniques for PPR computation and conduct a comprehensive survey of various recent PPR algorithms from an algorithmic perspective. We classify these approaches based on the types of queries they address and review their methodologies and contributions. We also discuss some representative algorithms for computing PPR on dynamic graphs and in parallel or distributed environments.

Group-fair recommendation aims at ensuring the equality of recommendation results across user groups categorized by sensitive attributes (e.g., gender, occupation, etc.). Existing group-fair recommendation models traditionally employ original user embeddings for both training and testing, primarily focusing on statistical learning while imposing group fairness constraints under the I.I.D. assumption. However, these models encounter limitations when addressing out-of-distribution (OOD) sensitive attributes. The fundamental issue of unfairness within user embeddings arises from a causal perspective, where each embedding vector comprises an exogenous component devoid of correlations with sensitive attributes and an endogenous component strongly correlated with these attributes. Overlooking the distinction between these two components during model training renders models sensitive to shifts in the distribution of sensitive attributes. This paper introduces the concept of Counterfactual Group Fairness (CGF) along with a corresponding metric to evaluate group fairness in scenarios involving OOD sensitive attributes in recommender systems. Building on this foundation, we propose a model-agnostic causal embedding learning framework named MACE. MACE effectively disentangles user embedding vectors into their exogenous and endogenous parts, thus ensuring group fairness, even in the presence of OOD sensitive attributes in embeddings. Specifically, MACE identifies the exogenous part of each user's embedding using mutual information minimization, treating it as instrumental variables. Subsequently, under the constraint of CGF, MACE reconstructs the endogenous and exogenous parts using the instrumental variable regression, combines the obtained parts into novel user embeddings using deep neural networks, and uses the combined embeddings for fair recommendation. Experimental results demonstrated that MACE can outperform the state-of-the-art baselines in terms of the metric of CGF while maintaining a comparable recommendation accuracy.

Document retrieval has been extensively studied within theindex-retrieveframework for decades, which has withstood the test of time. However, this approach inherently segregates the indexing and retrieval processes, preventing a cohesive, end-to-end optimization. To bridge this divide, we introduce WebUltron, a revolutionary model-centric indexer for document retrieval. This system embeds the entirety of document knowledge within the model, striving for seamless end-to-end retrieval. Two primary challenges with this indexer are the representation of document identifiers (docids) and the model's training. Current methods grapple with docids that lack semantic depth and the constraints of limited supervised data, making scaling up to larger datasets challenging. Addressing this, we've engineered two novel docid types imbued with richer semantics that also streamline model inference. Further enhancing WebUltron's capabilities, we've developed a three-stage training regimen, leveraging deeper corpus insights and fortifying query-docid relationships. Experiments on two public datasets demonstrate the superiority of WebUltron over advanced baselines for document retrieval.

Given a user, a personalized search model relies on her historical behaviors, such as issued queries and their clicked documents, to generate an interest profile and personalize search results accordingly. In interest profiling, most existing personalized search approaches use “static” document representations as the inputs, which do not change with the current search. However, a document is usually long and contains multiple pieces of information, a static fix-length document vector is usually insufficient to represent the important information related to the original query or the current query, and makes the profile noisy and ambiguous. To tackle this problem, we propose building dynamic and intent-oriented document representations which highlight important parts of a document rather than simply encode the entire text. Specifically, we divide each document into multiple passages, and then separately use the original query and the current query to interact with the passages. Thereafter we generate two “dynamic” document representations containing the key information around the historical and the current user intent, respectively. We then profile interest by capturing the interactions between these document representations, the historical queries, and the current query. Experimental results on a real-world search log dataset demonstrate that our model significantly outperforms state-of-the-art personalization methods.

Search result diversification plays a crucial role in improving users’ search experience by providing users with documents covering more subtopics. Previous studies have made great progress in leveraging inter-document interactions to measure the similarity among documents. However, different parts of the document may embody different subtopics and existing models ignore the subtle similarities and differences of content within each document. In this article, we propose a hierarchical attention framework to combine intra-document interactions with inter-document interactions in a complementary manner in order to conduct multi-grained document modeling. Specifically, we separate the document into passages to model the document content from multi-grained perspectives. Then, we design stacked interaction blocks to conduct inter-document and intra-document interactions. Moreover, to measure the subtopic coverage of each document more accurately, we propose a passage-aware document-subtopic interaction to perform fine-grained document-subtopic interaction. Experimental results demonstrate that our model achieves state-of-the-art performance compared with existing methods.

With the development of pre-trained language models, the dense retrieval models have become promising alternatives to the traditional retrieval models that rely on exact match and sparse bag-of-words representations. Different from most dense retrieval models using a bi-encoder to encode each query or document into a dense vector, the recently proposed late-interaction multi-vector models (i.e., ColBERT and COIL) achieve state-of-the-art retrieval effectiveness by using all token embeddings to represent documents and queries and modeling their relevance with a sum-of-max operation. However, these fine-grained representations may cause unacceptable storage overhead for practical search systems. In this study, we systematically analyze the matching mechanism of these late-interaction models and show that the sum-of-max operation heavily relies on the co-occurrence signals and some important words in the document. Based on these findings, we then propose several simple document pruning methods to reduce the storage overhead and compare the effectiveness of different pruning methods on different late-interaction models. We also leverage query pruning methods to further reduce the retrieval latency. We conduct extensive experiments on both in-domain and out-domain datasets and show that some of the used pruning methods can significantly improve the efficiency of these late-interaction models without substantially hurting their retrieval effectiveness.

Point-of-interest (POI) search is important for location-based services, such as navigation and online ride-hailing service. The goal of POI search is to find the most relevant destinations from a large-scale POI database given a text query. To improve the effectiveness and efficiency of POI search, most existing approaches are based on a multi-stage pipeline that consists of an efficiency-oriented retrieval stage and one or more effectiveness-oriented re-rank stages. In this article, we focus on the first efficiency-oriented retrieval stage of the POI search. We first identify the limitations of existing first-stage POI retrieval models in capturing the semantic-geography relationship and modeling the fine-grained geographical context information. Then, we propose a Geo-Enhanced Dense Retrieval framework for POI search to alleviate the above problems. Specifically, the proposed framework leverages the capacity of pre-trained language models (e.g., BERT) and designs a pre-training approach to better model the semantic match between the query prefix and POIs. With the POI collection, we first perform a token-level pre-training task based on a geographical-sensitive masked language prediction and design two retrieval-oriented pre-training tasks that link the address of each POI to its name and geo-location. With the user behavior logs collected from an online POI search system, we design two additional pre-training tasks based on users’ query reformulation behavior and the transitions between POIs. We also utilize a late-interaction network structure to model the fine-grained interactions between the text and geographical context information within an acceptable query latency. Extensive experiments on the real-world datasets collected from the Didichuxing application demonstrate that the proposed framework can achieve superior retrieval performance over existing first-stage POI retrieval methods.

Research on search result diversification strives to enhance the variety of subtopics within the list of search results. Existing studies usually treat a document as a whole and represent it with one fixed-length vector. However, considering that a long document could cover different aspects of a query, using a single vector to represent the document is usually insufficient. To tackle this problem, we propose to exploit multiple passages to better represent documents in search result diversification. Different passages of each document may reflect different subtopics of the query and comparison among the passages can improve result diversity. Specifically, we segment the entire document into multiple passages and train a classifier to filter out the irrelevant ones. Then the document diversity is measured based on several passages that can offer the information needs of the query. Thereafter, we devise a passage-aware search result diversification framework that takes into account the topic information contained in the selected document sequence and candidate documents. The candidate documents’ novelty is evaluated based on their passages while considering the dynamically selected document sequence. We conducted experiments on a commonly utilized dataset, and the results indicate that our proposed method performs better than the most leading methods.

Ambiguity in queries is a common problem in information retrieval. There are currently two solutions: search result personalization and diversification. The former aims to tailor results for different users based on their preferences, but the limitations are redundant results and incomplete capture of user intents. The goal of the latter is to return results that cover as many aspects related to the query as possible. It improves diversity yet loses personality and cannot return the exact results the user wants. Intuitively, such two solutions can complement each other and bring more satisfactory reranking results. In this article, we propose a novel framework, namely, PnD , to integrate personalization and diversification reasonably. We employ the degree of refinding to determine the weight of personalization dynamically. Moreover, to improve the diversity and relevance of reranked results simultaneously, we design a reset RNN structure (RRNN) with the “reset gate” to measure the influence of the newly selected document on novelty. Besides, we devise a “subtopic learning layer” to learn the virtual subtopics, which can yield fine-grained representations of queries, documents, and user profiles. Experimental results illustrate that our model can significantly outperform existing search result personalization and diversification methods.

Matching two natural language sentences is a fundamental problem in both natural language processing and information retrieval. Preliminary studies have shown that the syntactic structures help improve the matching accuracy, and different syntactic structures in natural language are complementary to sentence semantic understanding. Ideally, a matching model would leverage all syntactic information. Existing models, however, are only able to combine limited (usually one) types of syntactic information due to the complex and heterogeneous nature of the syntactic information. To deal with the problem, we propose a novel matching model, which formulates sentence matching as a representation learning task on a syntactic-informed heterogeneous graph. The model, referred to as SIGN (Syntactic-Informed Graph Network), first constructs a heterogeneous matching graph based on the multiple syntactic structures of two input sentences. Then the graph attention network algorithm is applied to the matching graph to learn the high-level representations of the nodes. With the help of the graph learning framework, the multiple syntactic structures, as well as the word semantics, can be represented and interacted in the matching graph and therefore collectively enhance the matching accuracy. We conducted comprehensive experiments on three public datasets. The results demonstrate that SIGN outperforms the state of the art and also can discriminate the sentences in an interpretable way.

Text retrieval is a long-standing research topic on information seeking, where a system is required to return relevant information resources to user’s queries in natural language. From heuristic-based retrieval methods to learning-based ranking functions, the underlying retrieval models have been continually evolved with the ever-lasting technical innovation. To design effective retrieval models, a key point lies in how to learn text representations and model the relevance matching. The recent success of pretrained language models (PLM) sheds light on developing more capable text-retrieval approaches by leveraging the excellent modeling capacity of PLMs. With powerful PLMs, we can effectively learn the semantic representations of queries and texts in the latent representation space, and further construct the semantic matching function between the dense vectors for relevance modeling. Such a retrieval approach is called dense retrieval , since it employs dense vectors to represent the texts. Considering the rapid progress on dense retrieval, this survey systematically reviews the recent progress on PLM-based dense retrieval. Different from previous surveys on dense retrieval, we take a new perspective to organize the related studies by four major aspects, including architecture, training, indexing and integration, and thoroughly summarize the mainstream techniques for each aspect. We extensively collect the recent advances on this topic, and include 300+ reference papers. To support our survey, we create a website for providing useful resources, and release a code repository for dense retrieval. This survey aims to provide a comprehensive, practical reference focused on the major progress for dense text retrieval.

In recent years, various dense retrieval methods have been developed to improve the performance of search engines with a vectorized index. However, these approaches require a large pre-computed index and have a limited capacity to memorize all semantics in a document within a single vector. To address these issues, researchers have explored end-to-end generative retrieval models that use a seq-to-seq generative model to directly return identifiers of relevant documents. Although these models have been effective, they are often trained with the MLE method. It only encourages the model to assign a high probability to the relevant document identifier, ignoring the relevance comparisons of other documents. This may lead to performance degradation in ranking tasks, where the core is to compare the relevance between documents. To address this issue, we propose a ranking-oriented generative retrieval model that incorporates relevance signals to better estimate the relative relevance of different documents in ranking tasks. Based upon the analysis of the optimization objectives of dense retrieval and generative retrieval, we propose utilizing dense retrieval to provide relevance feedback for generative retrieval. Under an alternate training framework, the generative retrieval model gradually acquires higher-quality ranking signals to optimize the model. Experimental results show that our approach increasing Recall@1 by 12.9% with respect to the baselines on MS MARCO dataset.

By scaling the model size, large pre-trained language models (PLMs) have shown remarkable performance in various natural language processing tasks, mostly outperforming small PLMs by a large margin.However, due to the high computational cost, the huge number of parameters also restricts the applicability of large PLMs in real-world systems.In this paper, we focus on scaling up the parameters of PLMs only during fine-tuning, to benefit from the over-parameterization, while without increasing the inference latency.Given a relatively small PLM, we over-parameterize it by employing a matrix product operator, an efficient and almost lossless decomposition method to factorize its contained parameter matrices into a set of higher-dimensional tensors.Considering the efficiency, we further propose both static and dynamic strategies to select the most important parameter matrices for over-parameterization.Extensive experiments have demonstrated that our approach can significantly boost the finetuning performance of small PLMs and even help small PLMs outperform 3× parameterized larger ones.

Although pre-trained language models (PLMs) have shown impressive performance by textonly self-supervised training, they are found lack of visual semantics or commonsense.Existing solutions often rely on explicit images for visual knowledge augmentation (requiring time-consuming retrieval or generation), and they also conduct the augmentation for the whole input text, without considering whether it is actually needed in specific inputs or tasks.To address these issues, we propose a novel Visually-Augmented fine-tuning approach that can be generally applied to various PLMs or NLP tasks, Without using any retrieved or generated Images, namely VAWI.Experimental results show that our approach can consistently improve the performance of BERT, RoBERTa, BART, and T5 at different scales, and outperform several competitive baselines on ten tasks.Our codes and data are publicly available at https://github.com/RUCAIBox/VAWI.

Recently, model-based retrieval has emerged as a new paradigm in text retrieval that discards the index in the traditional retrieval model and instead memorizes the candidate corpora using model parameters.This design employs a sequence-to-sequence paradigm to generate document identifiers, which enables the complete capture of the relevance between queries and documents and simplifies the classic indexretrieval-rerank pipeline.Despite its attractive qualities, there remain several major challenges in model-based retrieval, including the discrepancy between pre-training and fine-tuning, and the discrepancy between training and inference.To deal with the above challenges, we propose a novel two-stage model-based retrieval approach called TOME, which makes two major technical contributions, including the utilization of tokenized URLs as identifiers and the design of a two-stage generation architecture.We also propose a number of training strategies to deal with the training difficulty as the corpus size increases.Extensive experiments and analysis on MS MARCO and Natural Questions demonstrate the effectiveness of our proposed approach, and we investigate the scaling laws of TOME by examining various influencing factors.

People often imagine relevant scenes to aid in the writing process.In this work, we aim to utilize visual information for composition in the same manner as humans.We propose a method, LIVE, that makes pre-trained language models (PLMs) Learn to Imagine for Visuallyaugmented natural language gEneration.First, we imagine the scene based on the text: we use a diffusion model to synthesize high-quality images conditioned on the input texts.Second, we use CLIP to determine whether the text can evoke the imagination in a posterior way.Finally, our imagination is dynamic, and we conduct synthesis for each sentence rather than generate only one image for an entire paragraph.Technically, we propose a novel plug-and-play fusion layer to obtain visually-augmented representations for each text.Our vision-text fusion layer is compatible with Transformerbased architecture.We have conducted extensive experiments on four generation tasks using BART and T5, and the automatic results and human evaluation demonstrate the effectiveness of our proposed method.We will release the code, model, and data at the link: https://github.com/RUCAIBox/LIVE.

Within online platforms, it is critical to capture the dynamic user preference from the sequential interaction behaviors for making accurate recommendation over time. Recently, significant progress has been made in sequential recommendation with deep learning. However, existing neural sequential recommender often suffer from the data sparsity issue in real-world applications.

In online industrial advertising systems, conversion actions (e.g., purchases or downloads) often occur significantly delayed, even up to several days or weeks after the user clicks. This phenomenon leads to the crucial challenge calleddelayed feedback problem in streaming CVR prediction, that is, the online systems cannot receive the true label of conversions immediately for continuous training. To mitigate the delayed feedback problem, recent state-of-the-art methods often apply sample duplicate mechanisms to introduce early certain conversion information. Nevertheless, these works have overlooked a crucial issue of rapid shifts in data distribution and considered both the newly observed data and duplicated early data together, resulting in biases in both distributions. In this work, we propose a Dually enhanced Delayed Feedback Model (DDFM), which tackles the above issues by treating the newly observed data and duplicated early data separately. DDFM consists of dual unbiased CVR estimators that share the same form but utilize different latent variables as weights: one for the newly observed data and the other for the duplicated early data. To avoid high variance, we adopt an addition-only formula for these latent variables, eliminating multiplication or division operations. Furthermore, we design a shared-bottom network that efficiently and jointly estimates the latent variables in DDFM. Theoretical analysis demonstrates the unbiasedness and convergence properties of DDFM. Extensive experiments on both public and industrial large-scale real-world datasets exhibit that our proposed DDFM consistently outperforms existing state-of-the-art methods.

Document retrieval has always been a crucial problem in Web search. Recent works leverage pre-trained language models to represent documents in dense vectors. However, these works focus on the textual content but ignore the appearance of web pages (e.g., the visual style, the layout, and the images), which are actually essential for information delivery. To alleviate this problem, we propose a new dense retrieval model, namely VILE, to incorporate visual features into document representations. However, because a web page is usually very large and contains diverse information, simply concatenating its textual and visual features may result in a cluttered multi-modal representation that lacks focus on the important parts of the page. We observe that web pages often have a structured content organization, comprising multiple blocks that convey different information. Motivated by the observation, we propose building a multi-modal document representation by aggregating the fine-grained multi-modal block representations, to enable a more comprehensive understanding of the page. Specifically, we first segment a web page into multiple blocks, then create multi-modal features for each block. %allowing for more effective capture of its content and visual information. The representations of all blocks are then integrated into the final multi-modal page representation. VILE can better model the importance of different content regions, leading to a high-quality multi-modal representation. We collect screenshots and the corresponding layout information of some web pages in the MS MARCO Document Ranking dataset, resulting in a new multi-modal document retrieval dataset. Experimental results conducted on this dataset demonstrate that our model exhibits significant improvements over existing document retrieval models. Our code is available at https://github.com/yhy-2000/VILE.

The recent advent of end-to-end generative retrieval marks a significant shift in document retrieval methods, leveraging differentiable search indexes to directly produce relevant document identifiers (docids) in response to a specific query.Nevertheless, this approach faces two fundamental challenges: (i) a discrepancy between the token-level probabilistic optimization and the broader document-level relevance estimation; (ii) an overemphasis on top-1 results at the expense of overall ranking quality.To tackle these challenges, we propose a generative retrieval model with reinforcement learning from relevance feedback, which aims to align token-level docid generation with document-level relevance estimation.The training process incorporates three stages: supervised fine-tuning, relevance reward model training, and reinforced learning-to-rank from relevance feedback.To train a high-quality reward model, we define "relevance" under three progressive scenarios, which collectively offer a comprehensive evaluation of the document relevance.Experiments conducted on two benchmark datasets demonstrate the effectiveness of our proposed approach.

In this paper, we aim to improve the reasoning ability of large language models (LLMs) over structured data in a unified way.Inspired by the studies on tool augmentation for LLMs, we develop an Iterative Reading-then-Reasoning (IRR) framework to solve question answering tasks based on structured data, called StructGPT.In this framework, we construct the specialized interfaces to collect relevant evidence from structured data (i.e., reading), and let LLMs concentrate on the reasoning task based on the collected information (i.e., reasoning).Specially, we propose an invokinglinearization-generation procedure to support LLMs in reasoning on the structured data with the help of the interfaces.By iterating this procedure with provided interfaces, our approach can gradually approach the target answers to a given query.Experiments conducted on three types of structured data show that StructGPT greatly improves the performance of LLMs, under the few-shot and zero-shot settings.Our codes and data are publicly available at https://github.com/RUCAIBox/StructGPT.

Question Answering over Knowledge Graph (KGQA) aims to seek answer entities for the natural language question from a largescale Knowledge Graph (KG).To better perform reasoning on KG, recent work typically adopts a pre-trained language model (PLM) to model the question, and a graph neural network (GNN) based module to perform multihop reasoning on the KG.Despite the effectiveness, due to the divergence in model architecture, the PLM and GNN are not closely integrated, limiting the knowledge sharing and finegrained feature interactions.To solve it, we aim to simplify the above two-module approach, and develop a more capable PLM that can directly support subgraph reasoning for KGQA, namely ReasoningLM.In our approach, we propose a subgraph-aware self-attention mechanism to imitate the GNN for performing structured reasoning, and also adopt an adaptation tuning strategy to adapt the model parameters with 20,000 subgraphs with synthesized questions.After adaptation, the PLM can be parameter-efficient fine-tuned on downstream tasks.Experiments show that ReasoningLM surpasses state-of-the-art models by a large margin, even with fewer updated parameters and less training data.Our codes and data are publicly available at https://github.com/ RUCAIBox/ReasoningLM.

Large language models (LLMs), such as Chat-GPT, are prone to generate hallucinations, i.e., content that conflicts with the source or cannot be verified by the factual knowledge.To understand what types of content and to which extent LLMs are apt to hallucinate, we introduce the Hallucination Evaluation benchmark for Large Language Models (HaluEval), a large collection of generated and human-annotated hallucinated samples for evaluating the performance of LLMs in recognizing hallucination.To generate these samples automatically, we propose a two-stage framework, i.e., samplingthen-filtering.Besides, we hire some human labelers to annotate the hallucinations in Chat-GPT responses.The empirical results suggest that ChatGPT is likely to generate hallucinated content related to specific topics by fabricating unverifiable information (i.e., about 19.5% responses).Moreover, existing LLMs face great challenges in recognizing the hallucinations in texts.However, our experiments also prove that providing external knowledge or adding reasoning steps can help LLMs recognize hallucinations.Our benchmark can be accessed at https://github.com/RUCAIBox/HaluEval.

Inspired by the superior language abilities of large language models (LLM), large visionlanguage models (LVLM) have been recently proposed by integrating powerful LLMs for improving the performance on complex multimodal tasks.Despite the promising progress on LVLMs, we find that they suffer from object hallucinations, i.e., they tend to generate objects inconsistent with the target images in the descriptions.To investigate it, this work presents the first systematic study on object hallucination of LVLMs.We conduct the evaluation experiments on several representative LVLMs, and show that they mostly suffer from severe object hallucination issues.We further discuss that the visual instructions may influence the hallucination, and find that: objects that frequently appear in the visual instructions or co-occur with the image objects are obviously prone to be hallucinated by LVLMs.Besides, we further design a polling-based query method called POPE for better evaluation of object hallucination.Experiment results show that our POPE can evaluate object hallucination in a more stable and flexible way.

The recent success of large language models (LLMs) has shown great potential to develop more powerful conversational recommender systems (CRSs), which rely on natural language conversations to satisfy user needs.In this paper, we embark on an investigation into the utilization of ChatGPT for CRSs, revealing the inadequacy of the existing evaluation protocol.It might overemphasize the matching with ground-truth items annotated by humans while neglecting the interactive nature of CRSs.To overcome the limitation, we further propose an interactive Evaluation approach based on LLMs, named iEvaLM, which harnesses LLM-based user simulators.Our evaluation approach can simulate various system-user interaction scenarios.Through the experiments on two public CRS datasets, we demonstrate notable improvements compared to the prevailing evaluation protocol.Furthermore, we emphasize the evaluation of explainability, and Chat-GPT showcases persuasive explanation generation for its recommendations.Our study contributes to a deeper comprehension of the untapped potential of LLMs for CRSs and provides a more flexible and realistic evaluation approach for future research about LLMbased CRSs.The code is available at https: //github.com/RUCAIBox/iEvaLM-CRS.

The key of sequential recommendation lies in the accurate item correlation modeling. Previous models infer such information based on item co-occurrences, which may fail to capture the real causal relations, and impact the recommendation performance and explainability. In this paper, we equip sequential recommendation with a novel causal discovery module to capture causalities among user behaviors. Our general idea is firstly assuming a causal graph underlying item correlations, and then we learn the causal graph jointly with the sequential recommender model by fitting the real user behavior data. More specifically, in order to satisfy the causality requirement, the causal graph is regularized by a differentiable directed acyclic constraint. Considering that the number of items in recommender systems can be very large, we represent different items with a unified set of latent clusters, and the causal graph is defined on the cluster level, which enhances the model scalability and robustness. In addition, we provide theoretical analysis on the identifiability of the learned causal graph. To the best of our knowledge, this paper makes a first step towards combining sequential recommendation with causal discovery. For evaluating the recommendation performance, we implement our framework with different neural sequential architectures, and compare them with many state-of-the-art methods based on real-world datasets. Empirical studies manifest that our model can on average improve the performance by about 6.1% and 11.3% on F1and NDCG, respectively. To evaluate the model explainability, we build a new dataset with human labeled explanations for both quantitative and qualitative analysis.

Active learning is effective for tasks with limited labeled data by annotating a small set of data actively. It utilizes the current trained model to evaluate all unlabeled samples and annotates the best samples scored by a specific query strategy to update the underlying model iteratively. Most active learning approaches rely on only the current evaluation score but ignore the results from previous iterations. In this paper, we propose using more historical evaluation results which can provide additional information to help better select samples. First, we apply two heuristic features of the historical evaluation results, i.e. the weighted sum and the fluctuation of history sequences. Next, to make fuller use of the information contained in the historical results, we design a query strategy that learns to select samples based on the history sequence automatically. Our proposed idea is general and can be combined with both basic and state-of-the-art query strategies to achieve improvements. Experimental results show that our methods significantly promote existing methods.

Non-autoregressive (NAR) text generation has attracted much attention in the field of natural language processing, which greatly reduces the inference latency but has to sacrifice the generation accuracy. Recently, diffusion models, a class of latent variable generative models, have been introduced into NAR text generation, showing an improved text generation quality. In this survey, we review the recent progress in diffusion models for NAR text generation. As the background, we first present the general definition of diffusion models and the text diffusion models, and then discuss their merits for NAR generation. As the core content, we further introduce two mainstream diffusion models in existing work of text diffusion, and review the key designs of the diffusion process. Moreover, we discuss the utilization of pre-trained language models (PLMs) for text diffusion models and introduce optimization techniques for text data. Finally, we discuss several promising directions and conclude this paper. Our survey aims to provide researchers with a systematic reference of related research on text diffusion models for NAR generation. We also demonstrate our collection of text diffusion models at https://github.com/RUCAIBox/Awesome-Text-Diffusion-Models.

Providing personalization in product search has attracted increasing attention in both industry and research communities. Most existing personalized product search methods model users' individual search interests based on their historical search logs to generate personalized search results. However, the search logs may be sparse or noisy in the real scenario, which is difficult for existing methods to learn accurate and robust user representations. To address this issue, we propose a contrastive learning framework CoPPS that aims to learn high-quality user representations for personalized product search. Specifically, we design three data augmentation and contrastive learning strategies to construct self-supervision signals from the original search behaviours. The contrastive learning tasks utilize an external knowledge graph and exploit the correlations within and between user sequences, thereby facilitating the discovery of more meaningful search patterns and ultimately enhancing the quality of personalized search. Experimental results on the public Amazon datasets verify the effectiveness of our approach.

Recently, pretrained models have achieved remarkable performance not only in natural language processing but also in information retrieval (IR). Previous studies show that IR-oriented pretraining tasks can achieve better performance than only finetuning pretrained language models in IR datasets. Besides, the massive search log data obtained from mainstream search engines can be used in IR pretraining, for it contains users' implicit judgments of document relevance under a concrete query. However, existing methods mainly use direct query-document click signals to pretrain models. The potential supervision signals from search logs are far from being well explored. In this paper, we propose to comprehensively leverage four query-document relevance relations, including co-interaction and multi-hop relations, to pretrain ranking models in IR. Specifically, we focus on the user's click behavior and construct an Interaction Graph to represent the global relevance relations between queries and documents from all search logs. With the graph, we can consider the co-interaction and multi-hop q-d relationships through their neighbor nodes. Based on the relations extracted from the interaction graph, we propose four strategies to generate contrastive positive and negative q-d pairs and use these data to pretrain ranking models. Experimental results on both industrial and academic datasets demonstrate the effectiveness of our method.

Conversational recommender systems~(CRSs) aim to provide recommendation services via natural language conversations. Although a number of approaches have been proposed for developing capable CRSs, they typically rely on sufficient training data for training. Since it is difficult to annotate recommendation-oriented dialogue datasets, existing CRS approaches often suffer from the issue of insufficient training due to the scarcity of training data.

Although pre-trained language models~(PLMs) have recently advanced the research progress in mathematical reasoning, they are not specially designed as a capable multi-task solver, suffering from high cost for multi-task deployment (e.g. a model copy for a task) and inferior performance on complex mathematical problems in practical applications. To address these issues, we propose JiuZhang 2.0, a unified Chinese PLM specially for multi-task mathematical problem solving. Our idea is to maintain a moderate-sized model and employ the cross-task knowledge sharing to improve the model capacity in a multi-task setting. Specially, we construct a Mixture-of-Experts (MoE) architecture for modeling mathematical text, to capture the common mathematical knowledge across tasks. For optimizing the MoE architecture, we design multi-task continual pre-training and multi-task fine-tuning strategies for multi-task adaptation. These training strategies can effectively decompose the knowledge from the task data and establish the cross-task sharing via expert networks. To further improve the general capacity of solving different complex tasks, we leverage large language models (LLMs) as complementary models to iteratively refine the generated solution by our PLM, via in-context learning. Extensive experiments have demonstrated the effectiveness of our model.

Video moment localization aims to retrieve the target segment of an untrimmed video according to the natural language query. Weakly supervised methods gains attention recently, as the precise temporal location of the target segment is not always available. However, one of the greatest challenges encountered by the weakly supervised method is implied in the mismatch between the video and language induced by the coarse temporal annotations. To refine the vision-language alignment, recent works contrast the cross-modality similarities driven by reconstructing masked queries between positive and negative video proposals. However, the reconstruction may be influenced by the latent spurious correlation between the unmasked and the masked parts, which distorts the restoring process and further degrades the efficacy of contrastive learning since the masked words are not completely reconstructed from the cross-modality knowledge. In this paper, we discover and mitigate this spurious correlation through a novel proposed counterfactual cross-modality reasoning method. Specifically, we first formulate query reconstruction as an aggregated causal effect of cross-modality and query knowledge. Then by introducing counterfactual cross-modality knowledge into this aggregation, the spurious impact of the unmasked part contributing to the reconstruction is explicitly modeled. Finally, by suppressing the unimodal effect of masked query, we can rectify the reconstructions of video proposals to perform reasonable contrastive learning. Extensive experimental evaluations demonstrate the effectiveness of our proposed method. The code is available at https://github.com/sLdZ0306/CCR https://github.com/sLdZ0306/CCR.

Human motion prediction (HMP) has emerged as a popular research topic due to its diverse applications. Traditional methods rely on hand-crafted features and machine learning techniques, which often struggle to model the complex dynamics of human motion. Recent deep learning-based methods have achieved success by learning spatio-temporal representations of motion, but these models often overlook the reliability of motion data. Additionally, the temporal and spatial dependencies of skeleton nodes are distinct. The temporal relationship captures motion information over time, while the spatial relationship describes body structure and the relationships between different nodes. In this paper, we propose a novel spatio-temporal branching network using incremental information for HMP, which decouples the learning of temporal-domain and spatial-domain features, extracts more motion information, and achieves complementary cross-domain knowledge learning through knowledge distillation. Our approach effectively reduces noise interference and provides more expressive information for characterizing motion by separately extracting temporal and spatial features. We evaluate our approach on standard HMP benchmarks and outperform state-of-the-art methods in terms of prediction accuracy. Code is available at https://github.com/JasonWang959/STPMP.

Text-to-motion generation has gained increasing attention, but most existing methods are limited to generating short-term motions that correspond to a single sentence describing a single action. However, when a text stream describes a sequence of continuous motions, the generated motions corresponding to each sentence may not be coherently linked. Existing long-term motion generation methods face two main issues. Firstly, they cannot directly generate coherent motions and require additional operations such as interpolation to process the generated actions. Secondly, they generate subsequent actions in an autoregressive manner without considering the influence of future actions on previous ones. To address these issues, we propose a novel approach that utilizes a past-conditioned diffusion model with two optional coherent sampling methods: Past Inpainting Sampling and Compositional Transition Sampling. Past Inpainting Sampling completes subsequent motions by treating previous motions as conditions, while Compositional Transition Sampling models the distribution of the transition as the composition of two adjacent motions guided by different text prompts. Our experimental results demonstrate that our proposed method is capable of generating compositional and coherent long-term 3D human motions controlled by a user-instructed long text stream. The code is available at https://github.com/yangzhao1230/PCMDM

Contextual batched bandit (CBB) is a setting where a batch of rewards is observed from the environment at the end of each episode, but the rewards of the non-executed actions are unobserved, resulting in partial-information feedback. Existing approaches for CBB often ignore the rewards of the non-executed actions, leading to underutilization of feedback information. In this paper, we propose an efficient approach called Sketched Policy Updating with Imputed Rewards (SPUIR) that completes the unobserved rewards using sketching, which approximates the full-information feedbacks. We formulate reward imputation as an imputation regularized ridge regression problem that captures the feedback mechanisms of both executed and non-executed actions. To reduce time complexity, we solve the regression problem using randomized sketching. We prove that our approach achieves an instantaneous regret with controllable bias and smaller variance than approaches without reward imputation. Furthermore, our approach enjoys a sublinear regret bound against the optimal policy. We also present two extensions, a rate-scheduled version and a version for nonlinear rewards, making our approach more practical. Experimental results show that SPUIR outperforms state-of-the-art baselines on synthetic, public benchmark, and real-world datasets.

Explainable recommendation has attracted much attention from the industry and academic communities. It has shown great potential to improve the recommendation persuasiveness, informativeness and user satisfaction. In the past few years, while a lot of promising explainable recommender models have been proposed, the datasets used to evaluate them still suffer from several limitations, for example, the explanation ground truths are not labeled by the real users, the explanations are mostly single-modal and around only one aspect. To bridge these gaps, in this paper, we build a new explainable recommendation dataset, which, to our knowledge, is the first contribution that provides a large amount of real user labeled multi-modal and multi-aspect explaination ground truths. In specific, we firstly develop a video recommendation platform, where a series of questions around the recommendation explainability are carefully designed. Then, we recruit about 3000 high-quality labelers with different backgrounds to use the system, and collect their behaviors and feedback to our questions. In this paper, we detail the construction process of our dataset and also provide extensive analysis on its characteristics. In addition, we develop a library, where ten well-known explainable recommender models are implemented in a unified framework. Based on this library, we build several benchmarks for different explainable recommendation tasks. At last, we present many new opportunities brought by our dataset, which are expected to promote the field of explainable recommendation. Our dataset, library and the related documents have been released at https://reasoner2023.github.io/.

Chain-of-thought prompting (CoT) and tool augmentation have been validated in recent work as effective practices for improving large language models (LLMs) to perform step-by-step reasoning on complex math-related tasks.However, most existing math reasoning datasets may not be able to fully evaluate and analyze the ability of LLMs in manipulating tools and performing reasoning, as they often only require very few invocations of tools or miss annotations for evaluating intermediate reasoning steps, thus supporting only outcome evaluation.To address the issue, we construct CARP, a new Chinese dataset consisting of 4,886 computation-intensive algebra problems with formulated annotations on intermediate steps, facilitating the evaluation of the intermediate reasoning process.In CARP, we test four LLMs with CoT prompting, and find that they are all prone to make mistakes at the early steps of the solution, leading to incorrect answers.Based on this finding, we propose a new approach that can facilitate the deliberation on reasoning steps with tool interfaces, namely DELI.In DELI, we first initialize a step-by-step solution based on retrieved exemplars, then iterate two deliberation procedures that check and refine the intermediate steps of the generated solution, from both tool manipulation and natural language reasoning perspectives, until solutions converge or the maximum iteration is achieved.Experimental results on CARP and six other datasets show that the proposed DELI mostly outperforms competitive baselines, and can further boost the performance of existing CoT methods.Our data and code are available at https://github.com/RUCAIBox/CARP.

Next-basket Recommendation (NBR) refers to the task of predicting a set of items that a user will purchase in the next basket. However, most of existing works merely focus on the correlations between user preferences and predicted items, ignoring the essential correlations among items in the next basket, which often results in over-homogenization of predicted items. In this work, we presents a Generative next-basket Recommendation model (GenRec), a novel NBR paradigm that generates the recommended items one by one to form the next basket via an autoregressive decoder. This generative NBR paradigm contributes to capturing and considering item correlations inside each baskets in both training and serving. Moreover, we jointly consider user’s both item- and basket-level contextual information to better capture user’s multi-granularity preferences. Extensive experiments on three real-world datasets demonstrate the effectiveness of our model.

Conversational recommender systems (CRS) aim to provide the recommendation service via natural language conversations. To develop an effective CRS, high-quality CRS datasets are very crucial. However, existing CRS datasets suffer from the long-tail issue, i.e., a large proportion of items are rarely (or even never) mentioned in the conversations, which are called long-tail items. As a result, the CRSs trained on these datasets tend to recommend frequent items, and the diversity of the recommended items would be largely reduced, making users easier to get bored.

In the video recommendation, watch time is commonly adopted as an indicator of user interest. However, watch time is not only influenced by the matching of users’ interests but also by other factors, such as duration bias and noisy watching. Duration bias refers to the tendency for users to spend more time on videos with longer durations, regardless of their actual interest level. Noisy watching, on the other hand, describes users taking time to determine whether they like a video or not, which can result in users spending time watching videos they do not like. Consequently, the existence of duration bias and noisy watching make watch time an inadequate label for indicating user interest. Furthermore, current methods primarily address duration bias and ignore the impact of noisy watching, which may limit their effectiveness in uncovering user interest from watch time. In this study, we first analyze the generation mechanism of users’ watch time from a unified causal viewpoint. Specifically, we considered the watch time as a mixture of the user’s actual interest level, the duration-biased watch time, and the noisy watch time. To mitigate both the duration bias and noisy watching, we propose Debiased and Denoised watch time Correction (D2Co), which can be divided into two steps: First, we employ a duration-wise Gaussian Mixture Model plus frequency-weighted moving average for estimating the bias and noise terms; then we utilize a sensitivity-controlled correction function to separate the user interest from the watch time, which is robust to the estimation error of bias and noise terms. The experiments on two public video recommendation datasets and online A/B testing indicate the effectiveness of the proposed method.

Modern online service providers such as online shopping platforms often provide both search and recommendation (S&R) services to meet different user needs. Rarely has there been any effective means of incorporating user behavior data from both S&R services. Most existing approaches either simply treat S&R behaviors separately, or jointly optimize them by aggregating data from both services, ignoring the fact that user intents in S&R can be distinctively different. In our paper, we propose a Search-Enhanced framework for the Sequential Recommendation (SESRec) that leverages users' search interests for recommendation, by disentangling similar and dissimilar representations within S&R behaviors. Specifically, SESRec first aligns query and item embeddings based on users' query-item interactions for the computations of their similarities. Two transformer encoders are used to learn the contextual representations of S&R behaviors independently. Then a contrastive learning task is designed to supervise the disentanglement of similar and dissimilar representations from behavior sequences of S&R. Finally, we extract user interests by the attention mechanism from three perspectives, i.e., the contextual representations, the two separated behaviors containing similar and dissimilar interests. Extensive experiments on both industrial and public datasets demonstrate that SESRec consistently outperforms state-of-the-art models. Empirical studies further validate that SESRec successfully disentangle similar and dissimilar user interests from their S&R behaviors.

Legal case matching, which automatically constructs a model to estimate the similarities between the source and target cases, has played an essential role in intelligent legal systems. Semantic text matching models have been applied to the task where the source and target legal cases are considered as long-form text documents. These general-purpose matching models make the predictions solely based on the texts in the legal cases, overlooking the essential role of the law articles in legal case matching. In the real world, the matching results (e.g., relevance labels) are dramatically affected by the law articles because the contents and the judgments of a legal case are radically formed on the basis of law. From the causal sense, a matching decision is affected by the mediation effect from the cited law articles by the legal cases, and the direct effect of the key circumstances (e.g., detailed fact descriptions) in the legal cases. In light of the observation, this paper proposes a model-agnostic causal learning framework called Law-Match, under which the legal case matching models are learned by respecting the corresponding law articles. Given a pair of legal cases and the related law articles, Law-Match considers the embeddings of the law articles as instrumental variables(IVs), and the embeddings of legal cases as treatments. Using IV regression, the treatments can be decomposed into law-related and law-unrelated parts, respectively reflecting the mediation and direct effects. These two parts are then combined with different weights to collectively support the final matching prediction. We show that the framework is model-agnostic, and a number of legal case matching models can be applied as the underlying models. Comprehensive experiments show that Law-Match can outperform state-of-the-art baselines on three public datasets.

Learning effective high-order feature interactions is very crucial in the CTR prediction task. However, it is very time-consuming to calculate high-order feature interactions with massive features in online e-commerce platforms. Most existing methods manually design a maximal order and further filter out the useless interactions from them. Although they reduce the high computational costs caused by the exponential growth of high-order feature combinations, they still suffer from the degradation of model capability due to the suboptimal learning of the restricted feature orders. The solution to maintain the model capability and meanwhile keep it efficient is a technical challenge, which has not been adequately addressed. To address this issue, we propose an adaptive feature interaction learning model, named as EulerNet, in which the feature interactions are learned in a complex vector space by conducting space mapping according to Euler's formula. EulerNet converts the exponential powers of feature interactions into simple linear combinations of the modulus and phase of the complex features, making it possible to adaptively learn the high-order feature interactions in an efficient way. Furthermore, EulerNet incorporates the implicit and explicit feature interactions into a unified architecture, which achieves the mutual enhancement and largely boosts the model capabilities. Such a network can be fully learned from data, with no need of pre-designed form or order for feature interactions. Extensive experiments conducted on three public datasets have demonstrated the effectiveness and efficiency of our approach. Our code is available at: https://github.com/RUCAIBox/EulerNet.

In recent years, the reproducibility of recommendation models has become a severe concern in recommender systems. In light of this challenge, we have previously released a unified, comprehensive and efficient recommendation library called RecBole, attracting much attention from the research community. With the increasing number of users, we have received a number of suggestions and update requests. This motivates us to make further improvements on our library, so as to meet the user requirements and contribute to the research community. In this paper, we present a significant update of RecBole, making it more user-friendly and easy-to-use as a comprehensive benchmark library for recommendation. More specifically, the highlights of this update are summarized as: (1) we include more benchmark models and datasets, improve the benchmark framework in terms of data processing, training and evaluation, and release reproducible configurations to benchmark the recommendation models; (2) we upgrade the user friendliness of our library by providing more detailed documentation and well-organized frequently asked questions, and (3) we propose several development guidelines for the open-source library developers. These extensions make it much easier to reproduce the benchmark results and stay up-to-date with the recent advances on recommender systems. Our update is released at the link: https://github.com/RUCAIBox/RecBole.

With the growth of high-dimensional sparse data in web-scale recommender systems, the computational cost to learn high-order feature interaction in CTR prediction task largely increases, which limits the use of high-order interaction models in real industrial applications. Some recent knowledge distillation based methods transfer knowledge from complex teacher models to shallow student models for accelerating the online model inference. However, they suffer from the degradation of model accuracy in knowledge distillation process. It is challenging to balance the efficiency and effectiveness of the shallow student models. To address this problem, we propose a Directed Acyclic Graph Factorization Machine (KD-DAGFM) to learn the high-order feature interactions from existing complex interaction models for CTR prediction via Knowledge Distillation. The proposed lightweight student model DAGFM can learn arbitrary explicit feature interactions from teacher networks, which achieves approximately lossless performance and is proved by a dynamic programming algorithm. Besides, an improved general model KD-DAGFM+ is shown to be effective in distilling both explicit and implicit feature interactions from any complex teacher model. Extensive experiments are conducted on four real-world datasets, including a large-scale industrial dataset from WeChat platform with billions of feature dimensions. KD-DAGFM achieves the best performance with less than 21.5% FLOPs of the state-of-the-art method on both online and offline experiments, showing the superiority of DAGFM to deal with the industrial scale data in CTR prediction task.

Alleviating the examination and trust bias in ranking systems is an important research line in unbiased learning-to-rank (ULTR). Current methods typically use the propensity to correct the biased user clicks and then learn ranking models based on the corrected clicks. Though successes have been achieved, directly modifying the clicks suffers from the inherent high variance because the propensities are usually involved in the denominators of corrected clicks. The problem gets even worse in the situation of mixed examination and trust bias. To address the issue, this paper proposes a novel ULTR method called Decomposed Ranking Debiasing (DRD). DRD is tailored for learning unbiased relevance models with low variance in the existence of examination and trust bias. Unlike existing methods that directly modify the original user clicks, DRD proposes to decompose each click prediction as the combination of a relevance term outputted by the ranking model and other bias terms. The unbiased relevance model, therefore, can be learned by fitting the overall click predictions to the biased user clicks. A joint learning algorithm is developed to learn the relevance and bias models' parameters alternatively. Theoretical analysis showed that, compared with existing methods, DRD has lower variance while retains unbiasedness. Empirical studies indicated that DRD can effectively reduce the variance and outperform the state-of-the-art ULTR baselines.

The key to personalized search is modeling user intents to tailor returned results for different users. Existing personalized methods mainly focus on learning implicit user interest vectors. In this paper, we propose ExpliPS, a personalized search model that explicitly incorporates query subtopics into personalization. It models the user’s current intent by estimating the user’s preference over the subtopics of the current query and personalizes the results over the weighted subtopics. We think that in such a way, personalized search could be more explainable and stable. Specifically, we first employ a semantic encoder to learn the representations of the user’s historical behaviours. Then with the historical behaviour representations, a subtopic preference encoder is devised to predict the user’s subtopic preferences on the current query. Finally, we rerank the candidates via a subtopic-aware ranker that prioritizes the documents relevant to the user-preferred subtopics. Experimental results show our model ExpliPS outperforms the state-of-the-art personalized web search models with explainable and stable results.

Recently, contrastive learning has been shown effective in fine-tuning pre-trained language models (PLM) to learn sentence representations, which incorporates perturbations into unlabeled sentences to augment semantically related positive examples for training. However, previous works mostly adopt heuristic perturbation methods that are independent of the sentence representations. Since the perturbations are unaware of the goal or process of sentence representation learning during training, it is likely to lead to sub-optimal augmentations for conducting constrative learning. To address this issue, we propose a new frameworkL2P-CSRthat adopts a learnable perturbation strategy for improving contrastive learning of sentence representations. In our L2P-CSR, we design a safer perturbation mechanism that only weakens the influence of tokens and features on the sentence representation, which avoids dramatically changing the semantics of the sentence representations. Besides, we devise a gradient-based algorithm to generate adaptive perturbations specially for the dynamically updated sentence representation during training. Such a way is more capable of augmenting high-quality examples that guide the sentence representation learning. Extensive experiments on diverse sentence-related tasks show that our approach outperforms competitive baselines.

Deep semantic matching aims at discriminating the relationship between documents based on deep neural networks. In recent years, it becomes increasingly popular to organize documents with a graph structure, then leverage both the intrinsic document features and the extrinsic neighbor features to derive discrimination. Most of the existing works mainly care about how to utilize the presented neighbors, whereas limited effort is made to filter appropriate neighbors. We argue that the neighbor features could be highly noisy and partially useful. Thus, a lack of effective neighbor selection will not only incur a great deal of unnecessary computation cost but also restrict the matching accuracy severely. In this work, we propose a novel framework, C ascaded D eep S emantic M atching ( CDSM ), for accurate and efficient semantic matching on textual graphs. CDSM is highlighted for its two-stage workflow. In the first stage, a lightweight CNN-based ad-hod neighbor selector is deployed to filter useful neighbors for the matching task with a small computation cost. We design both one-step and multi-step selection methods. In the second stage, a high-capacity graph-based matching network is employed to compute fine-grained relevance scores based on the well-selected neighbors. It is worth noting that CDSM is a generic framework which accommodates most of the mainstream graph-based semantic matching networks. The major challenge is how the selector can learn to discriminate the neighbors’ usefulness which has no explicit labels. To cope with this problem, we design a weak-supervision strategy for optimization, where we train the graph-based matching network at first and then the ad-hoc neighbor selector is learned on top of the annotations from the matching network. We conduct extensive experiments with three large-scale datasets, showing that CDSM notably improves the semantic matching accuracy and efficiency thanks to the selection of high-quality neighbors. The source code is released at https://github.com/jingjyyao/CDSM.

Sequential recommendation has recently attracted increasing attention from the industry and academic communities. While previous models have achieved remarkable successes, an important problem may still hinder their performances, that is, the sparsity of the real-world data. In this paper, we propose a novel counterfactual data augmentation framework to alleviate the problem of data sparsity. In specific, our framework contains a sampler model and an anchor model. The sampler model aims to generate high-quality user behavior sequences, while the anchor model is trained based on the original and new generated samples, and leveraged to provide the final recommendation list. To implement the sampler model, we first design four types of heuristic methods based on either random or frequency-based strategies. And then, to improve the quality of the generated sequences, we propose two learning-based samplers by discovering the decision boundaries or increasing the sample informativeness. At last, we build an RL based model to automatically determine where to edit the history behaviors and how many items should be replaced. Considering that the sampler model can be imperfect, we, at last, analyze the influence of the noisy information contained in the generated sequences on the anchor model in theory, and design a simple but effective method to better serve the anchor model. We conduct extensive experiments to demonstrate the effectiveness of our model.

Knowledge base question answering (KBQA) aims to answer a question over a knowledge base (KB). Early studies mainly focused on answering simple questions over KBs and achieved great success. However, their performances on complex questions are still far from satisfactory. Therefore, in recent years, researchers propose a large number of novel methods, which looked into the challenges of answering complex questions. In this survey, we review recent advances in KBQA with the focus on solving complex questions, which usually contain multiple subjects, express compound relations, or involve numerical operations. In detail, we begin with introducing the complex KBQA task and relevant background. Then, we present two mainstream categories of methods for complex KBQA, namely semantic parsing-based (SP-based) methods and information retrieval-based (IR-based) methods. Specifically, we illustrate their procedures with flow designs and discuss their difference and similarity. Next, we summarize the challenges that these two categories of methods encounter when answering complex questions, and explicate advanced solutions as well as techniques used in existing work. After that, we discuss the potential impact of pre-trained language models (PLMs) on complex KBQA. To help readers catch up with SOTA methods, we also provide a comprehensive evaluation and resource about complex KBQA task. Finally, we conclude and discuss several promising directions related to complex KBQA for future research.

While self-supervised learning techniques are often used to mine hidden knowledge from unlabeled data via modeling multiple views, it is unclear how to perform effective representation learning in a complex and inconsistent context. To this end, we propose a new multi-view self-supervised learning method, namelyconsistency and complementarity network(CoCoNet), to comprehensively learn global inter-view consistent and local cross-view complementarity-preserving representations from multiple views. To capture crucial common knowledge which is implicitly shared among views, CoCoNet employs a global consistency module that aligns the probabilistic distribution of views by utilizing an efficient discrepancy metric based on the generalized sliced Wasserstein distance. To incorporate cross-view complementary information, CoCoNet proposes a heuristic complementarity-aware contrastive learning approach, which extracts a complementarity-factor jointing cross-view discriminative knowledge and uses it as the contrast to guide the learning of view-specific encoders. Theoretically, the superiority of CoCoNet is verified by our information-theoretical-based analyses. Empirically, our thorough experimental results show that CoCoNet outperforms the state-of-the-art self-supervised methods by a significant margin, for instance, CoCoNet beats the best benchmark method by an average margin of 1.1% on ImageNet.

The goal of personalized search is to tailor the document ranking list to meet user's individual needs. Previous studies showed users usually look for the information that has been searched before. This is called re-finding behavior which is widely explored in existing personalized search approaches. However, most existing methods for identifying re-finding behavior focus on simple lexical similarities between queries. In this paper, we propose a personalized framework based on hierarchical memory networks (MN) to enhance the identification of the potential re-finding behavior. Specifically, we explore the potential re-finding behaviors of users from two dimensions. (1) Granularity dimension. The framework carries out re-finding identification with external memories from word, sentence, and session levels. (2) Query intent dimension. Query-based re-finding and document-based re-finding are taken into account to cover user's different query intents. To enhance the interaction between different memory slots, we optimize the$READ$operation of MN with two strategies that utilize the information in memory in a multi-hop way. Endowed with these memory networks, we can enhance user's potential re-finding behaviors and build a fine-grained user model dynamically. Experimental results on two datasets have a significant improvement over baselines, and the optimized$READ$operation shows better performance.

Recent studies show that historical behaviors (such as queries and their clicks) contained in a search session can benefit the ranking performance of subsequent queries in the session. Existing neural context-aware ranking models usually rank documents based on either latent representations of user search behaviors or the word-level interactions between the candidate document and each historical behavior in the search session. However, these two kinds of models both have their own drawbacks. Representation-based models neglect fine-grained information on word-level interactions, whereas interaction-based models suffer from the length restriction of session sequence because of the large cost of word-level interactions. To complement the limitations of these two kinds of models, we propose a unified context-aware document ranking model that takes full advantage of both representation and interaction. Specifically, instead of matching a candidate document with every single historical query in a session, we encode the session history into a latent representation and use this representation to enhance the current query and the candidate document. We then just match the enhanced query and candidate document with several matching components to capture the fine-grained information of word-level interactions. Rich experiments on two public query logs prove the effectiveness and efficiency of our model for leveraging representation and interaction.

Search result diversification aims to generate diversified search results so as to meet the various information needs of users. Most of those existing diversification methods greedily select the optimal documents one-by-one comparing with the selected document sequences. Due to the fact that the information utilities of the candidate documents are not independent, a model based on greedy document selection may not produce the global optimal ranking results. To address this issue, some work proposes to model global document interactions regardless of whether a document is selected, which is inconsistent with actual user behavior. In this article, we propose a new supervised diversification framework as an ensemble of global interaction and document selection. Based on a self-attention encoder-decoder structure and an RNN-based document selection component, the model can simultaneously leverage both the global interactions among all the documents and the interactions between the selected sequence and each unselected document. This framework is called Greedy Diversity Encoder with Self-Attention (GDESA). Experimental results show that GDESA outperforms previous methods that rely just on global interactions, and our further analysis demonstrates that using both global interactions and document selection is necessary and beneficial.

Recommender systems are currently widely used in various applications helping people filter information. Existing models always embed the rich information for recommendation, such as items, users, and contexts in real-value vectors, and make predictions based on these vectors. In the view of causal inference, the associations between representation vectors and user feedback are inevitably a mixture of the causal part that describes why a user prefers an item, and the non-causal part that merely reflects the statistical dependencies, for example, the display ranking position and sales promotion. However, most recommender systems assume the user-item interactions are only affected by user preferences, neglecting the striking differences between these two associations. To address this problem, we propose a model-agnostic causal learning framework called IV4Rec+ that can effectively decompose the embedding vectors into these two parts. Moreover, two strategies are proposed to utilize search queries as instrumental variables: IV4Rec+(I) only decomposes the item embeddings, while IV4Rec+(UI) decomposes both user and item embeddings. IV4Rec+ is a model-agnostic design that can be applied to many existing recommender systems, e.g., DIN, NRHUB, and SRGNN. Extensive experiments on three datasets show that IV4Rec+ significantly facilitates the performance of recommender systems and outperforms state-of-the-art frameworks.

To characterize complex and heterogeneous side information in recommender systems, the heterogeneous information network (HIN) has shown superior performance and attracted much research attention. In HIN, the rich entities, relations, and paths can be utilized to model the correlations of users and items; such a task setting is often calledHIN-based recommendation. Although HIN provides a general approach to modeling rich side information, it lacks special consideration on the goal of the recommendation task. The aggregated context from the heterogeneous graph is likely to incorporate irrelevant information, and the learned representations are not specifically optimized according to the recommendation task. Therefore, there is a need to rethink how to leverage the useful information from HIN to accomplish the recommendation task. To address the above issues, we propose a Curriculum pre-training based HEterogeneous Subgraph Transformer (calledCHEST) with newdata characterization,representation model,andlearning algorithm. Specifically, we consider extracting useful information from HIN to compose the interaction-specific heterogeneous subgraph, containing highly relevant context information for recommendation. Then, we capture the rich semantics (e.g., graph structure and path semantics) within the subgraph via a heterogeneous subgraph Transformer, where we encode the subgraph into multi-slot sequence representations. Besides, we design a curriculum pre-training strategy to provide an elementary-to-advanced learning process. The elementary course focuses on capturing local context information within the subgraph, and the advanced course aims to learn global context information. In this way, we gradually capture useful semantic information from HIN for modeling user-item interactions. Extensive experiments conducted on four real-world datasets demonstrate the superiority of our proposed method over a number of competitive baselines, especially when only limited training data is available.

Legal judgment prediction (LJP) is a fundamental task of legal artificial intelligence. It aims to automatically predict the judgment results of legal cases. Three typical subtasks are relevant law article prediction, charge prediction, and term-of-penalty prediction. Due to the wide range of potential applications, LJP has attracted a great deal of interest, prompting the development of numerous approaches. These methods mainly focus on building a more accurate representation of a case’s fact description in order to improve the performance of judgment prediction. They overlook, however, the practical judicial scenario in which human judges often compare similar law articles or possible charges before making a final decision. To this end, we propose a supervised contrastive learning framework for the LJP task. Specifically, we train the model to distinguish (1) various law articles within the same chapter of a Law and (2) similar charges of the same law article or related law articles. By this means, the fine-grained differences between similar articles/charges can be captured, which are important for making a judgment. Besides, we optimize our model by identifying cases with the same article/charge labels, allowing it to more effectively model the relationship between the case’s fact description and its associated labels. By jointly learning the LJP task with the aforementioned contrastive learning tasks, our model achieves better performance than the state-of-the-art models on two real-world datasets.

In recommender systems, top- N recommendation is an important task with implicit feedback data. Although the recent success of deep learning largely pushes forward the research on top- N recommendation, there are increasing concerns on appropriate evaluation of recommendation algorithms. It therefore is important to study how recommendation algorithms can be reliably evaluated and thoroughly verified. This work presents a large-scale, systematic study on six important factors from three aspects for evaluating recommender systems. We carefully select 12 top- N recommendation algorithms and eight recommendation datasets. Our experiments are carefully designed and extensively conducted with these algorithms and datasets. In particular, all the experiments in our work are implemented based on an open sourced recommendation library, Recbole [ 139 ], which ensures the reproducibility and reliability of our results. Based on the large-scale experiments and detailed analysis, we derive several key findings on the experimental settings for evaluating recommender systems. Our findings show that some settings can lead to substantial or significant differences in performance ranking of the compared algorithms. In response to recent evaluation concerns, we also provide several suggested settings that are specially important for performance comparison.

Sequential recommendation models aim to predict the interested items to a user based on his historical behaviors. To train sequential recommenders, implicit feedback data is widely adopted since it is easier to obtain than explicit feedback data. In the setting of implicit feedback, a user’s historical behaviors can be characterized as a chronologically ordered sequence of interacted items. From a perspective of machine learning, the historical interaction sequence and the recommended items can be considered as context and label , respectively, which are usually in one-hot representations in the recommendation models. However, due to the discrete nature, one-hot representations are hard to sufficiently reflect the underlying user preference, and might also contain noise from implicit feedback that will mislead the model training. To solve these issues, we propose a general optimization framework, Multi-View Smoothness (MVS), to enhance the smoothness of sequential recommendation models in both data representations and model learning. Specifically, with the help of a complementary model, we smooth and enrich the one-hot representations of contexts and labels to better depict the underlying user preference (i.e., context smoothness and label smoothness), and devise a model regularization strategy to enforce the neighborhood smoothness of the model itself (i.e., model smoothness). Based on these strategies, we design three regularizers to constrain and improve the training of sequential recommendation models. Extensive experiments on five datasets show that our approach is able to improve the performance of various base models consistently and outperform other regularization training methods.

Knowledge-grounded conversation (KGC) shows great potential in building an engaging and knowledgeable chatbot, and knowledge selection is a key ingredient in it.However, previous methods for knowledge selection only concentrate on the relevance between knowledge and dialogue context, ignoring the fact that age, hobby, education and life experience of an interlocutor have a major effect on his or her personal preference over external knowledge.Without taking the personalization issue into account, it is difficult to select the proper knowledge and generate persona-consistent responses.In this work, we introduce personal memory into knowledge selection in KGC to address the personalization issue.We propose a variational method to model the underlying relationship between one's personal memory and his or her selection of knowledge, and devise a learning scheme in which the forward mapping from personal memory to knowledge and its inverse mapping is included in a closed loop so that they could teach each other.Experiment results show that our method outperforms existing KGC methods significantly on both automatic evaluation and human evaluation.

In this paper, we study how to continually pretrain language models for improving the understanding of math problems.Specifically, we focus on solving a fundamental challenge in modeling math problems, i.e., how to fuse the semantics of textual description and formulas, which are highly different in essence.To address this issue, we propose a new approach called COMUS to continually pre-train language models for math problem understanding with syntax-aware memory network.In this approach, we first construct the math syntax graph to model the structural semantic information, by combining the parsing trees of the text and formulas, and then design the syntax-aware memory networks to deeply fuse the features from the graph and text.With the help of syntax relations, we can model the interaction between the token from the text and its semantic-related nodes within the formulas, which is helpful to capture fine-grained semantic correlations between texts and formulas.Besides, we devise three continual pre-training tasks to further align and fuse the representations of the text and math syntax graph.Experimental results on four tasks in the math domain demonstrate the effectiveness of our approach.Our code and data are publicly available at the link: https: //github.com/RUCAIBox/COMUS.

Applying existing methods to emotional support conversation-which provides valuable assistance to people who are in need-has two major limitations: (a) they generally employ a conversation-level emotion label, which is too coarse-grained to capture user's instant mental state; (b) most of them focus on expressing empathy in the response(s) rather than gradually reducing user's distress.To address the problems, we propose a novel model MISC, which firstly infers the user's fine-grained emotional status, and then responds skillfully using a mixture of strategy.Experimental results on the benchmark dataset demonstrate the effectiveness of our method and reveal the benefits of fine-grained emotion understanding as well as mixed-up strategy modeling.Our code and data could be found in https: //github.com/morecry/MISC.

Recently, contrastive learning has been shown to be effective in improving pre-trained language models (PLM) to derive high-quality sentence representations. It aims to pull close positive examples to enhance the alignment while push apart irrelevant negatives for the uniformity of the whole representation space.However, previous works mostly adopt in-batch negatives or sample from training data at random. Such a way may cause the sampling bias that improper negatives (false negatives and anisotropy representations) are used to learn sentence representations, which will hurt the uniformity of the representation space.To address it, we present a new framework DCLR (Debiased Contrastive Learning of unsupervised sentence Representations) to alleviate the influence of these improper negatives.In DCLR, we design an instance weighting method to punish false negatives and generate noise-based negatives to guarantee the uniformity of the representation space.Experiments on seven semantic textual similarity tasks show that our approach is more effective than competitive baselines. Our code and data are publicly available at the link: bluehttps://github.com/RUCAIBox/DCLR.

Within online platforms, it is critical to capture the semantics of sequential user behaviors for accurately predicting user interests. Recently, significant progress has been made in sequential recommendation with deep learning. However, existing neural sequential recommendation models may not perform well in practice due to the sparsity of the real-world data especially in cold-start scenarios. To tackle this problem, we propose the model ReDA, which stands for Retrieval-enhanced Data Augmentation for modeling sequential user behaviors. The main idea of our approach is to leverage the related information from similar users for generating both relevant and diverse augmentation. First, we train a neural retriever to retrieve the augmentation users according to the se- mantic similarity between user representations, and then conduct two types of data augmentation to generate augmented user representations. Furthermore, these augmented data are incorporated in a contrastive learning framework for learning more capable representations. Extensive experiments conducted on both public and industry datasets demonstrate the superiority of our proposed method over existing state-of-the-art methods, especially when only limited training data is available.

Users' search tasks have become increasingly complicated, requiring multiple queries and interactions with the results. Recent studies have demonstrated that modeling the historical user behaviors in a session can help understand the current search intent. Existing context-aware ranking models primarily encode the current session sequence (from the first behavior to the current query) and compute the ranking score using the high-level representations. However, there is usually some noise in the current session sequence (useless behaviors for inferring the search intent) that may affect the quality of the encoded representations. To help the encoding of the current user behavior sequence, we propose to use a decoder and the information of future sequences and a supplemental query. Specifically, we design three generative tasks that can help the encoder to infer the actual search intent: (1) predicting future queries, (2) predicting future clicked documents, and (3) predicting a supplemental query. We jointly learn the ranking task with these generative tasks using an encoder-decoder structured approach. Extensive experiments on two public search logs demonstrate that our model outperforms all existing baselines, and the designed generative tasks can actually help the ranking task. Besides, additional experiments also show that our approach can be easily applied to various Transformer-based encoder-decoder models and improve their performance.

Sequential recommender systems train their models based on a large amount of implicit user feedback data and may be subject to biases when users are systematically under/over-exposed to certain items. Unbiased learning based on inverse propensity scores (IPS), which estimate the probability of observing a user-item pair given the historical information, has been proposed to address the issue. In these methods, propensity score estimation is usually limited to the view of item, that is, treating the feedback data as sequences of items that interacted with the users. However, the feedback data can also be treated from the view of user, as the sequences of users that interact with the items. Moreover, the two views can jointly enhance the propensity score estimation. Inspired by the observation, we propose to estimate the propensity scores from the views of user and item, called Dually Enhanced Propensity Score Estimation (DEPS). Specifically, given a target user-item pair and the corresponding item and user interaction sequences, DEPS first constructs a time-aware causal graph to represent the user-item observational probability. According to the graph, two complementary propensity scores are estimated from the views of item and user, respectively, based on the same set of user feedback data. Finally, two transformers are designed to make use of the two propensity scores and make the final preference prediction. Theoretical analysis showed the unbiasedness and variance of DEPS. Experimental results on three publicly available benchmarks and a proprietary industrial dataset demonstrated that DEPS can significantly outperform the state-of-the-art baselines.

In order to support the study of recent advances in recommender systems, this paper presents an extended recommendation library consisting of eight packages for up-to-date topics and architectures. First of all, from a data perspective, we consider three important topics related to data issues (ie sparsity, bias and distribution shift ), and develop five packages accordingly, including meta-learning, data augmentation, debiasing, fairness and cross-domain recommendation. Furthermore, from a model perspective, we develop two benchmarking packages for Transformer-based and graph neural network~(GNN)-based models, respectively. All the packages (consisting of 65 new models) are developed based on a popular recommendation framework RecBole, ensuring that both the implementation and interface are unified. For each package, we provide complete implementations from data loading, experimental setup, evaluation and algorithm implementation. This library provides a valuable resource to facilitate the up-to-date research in recommender systems. The project is released at the link: \urlhttps://github.com/RUCAIBox/RecBole2.0.

Recently, Mixture-of-Experts (short as MoE) architecture has achieved remarkable success in increasing the model capacity of large-scale language models. However, MoE requires incorporating significantly more parameters than the base model being extended. In this paper, we propose building a parameter-efficient MoE architecture by sharing information across experts. We adopt matrix product operator (MPO, a tensor decomposition from quantum many-body physics) to reconstruct the parameter matrix in the expert layer and increase model capacity for pre-trained language models by sharing parameters of the central tensor (containing the core information) among different experts while enabling the specificity through the auxiliary tensors (complementing the central tensor) of different experts. To address the unbalanced optimization issue, we further design the gradient mask strategy for the MPO-based MoE architecture. Extensive experiments based on T5 and GPT-2 show improved performance and efficiency of the pre-trained language model (27.2x reduction in total parameters for the superior model performance, compared with the Switch Transformers). Our code is publicly available at https://github.com/RUCAIBox/MPO/MPOE.

Recently, pretrained language models (PLMs) have had exceptional success in language generation. To leverage the rich knowledge encoded by PLMs, a simple yet powerful paradigm is to use prompts in the form of either discrete tokens or continuous embeddings. In existing studies, these prompting methods are typically independent of the inputs, lacking sufficient consideration of input semantics. To address this issue, we propose a novel continuous prompting approach, called context-tuning, to fine-tuning PLMs for natural language generation. Firstly, the prompts are derived based on the input text to elicit useful knowledge from PLMs for generation. We refer to such prompts as contextualized prompts. Secondly, we use continuous inverse prompting to improve the process of natural language generation by modeling an inverse generation process from output to input, making the generated text more relevant to the inputs. Furthermore, we utilize a lightweight context-tuning method that fine-tunes only 0.12% of the parameters while maintaining good performance. Our code is publicly available at https://github.com/RUCAIBox/Context-Tuning.

Studies have shown that the sentence’s syntactic structures are important for semantic sentence matching. A typical approach is encoding each sentence’s syntactic structure into an embedding vector, which can be combined with other features to predict the final matching scores. Though successes have been observed, embedding the whole syntactic structures as one vector inevitably overlooks the fine-grained syntax matching patterns, e.g. the alignment of specific term dependencies relations in the two inputted sentences. In this paper, we formalize the task of semantic sentence matching as a problem of graph matching in which each sentence is represented as a directed graph according to its syntactic structures. The syntax matching patterns (i.e. similar syntactic structures) between two sentences, therefore, can be extracted as the sub-graph structure alignments. The proposed method, referred to as Interacted Syntax Graphs (ISG), represents two sentences’ syntactic alignments as well as their semantic matching signals into one association graph. After that, the neural quadratic assignment programming (QAP) is adapted to extract syntactic matching patterns from the association graph. In this way, the syntactic structures fully interact in a fine granularity during the matching process. Experimental results on three public datasets demonstrated that ISG can outperform the state-of-the-art baselines effectively and efficiently. The empirical analysis also showed that ISG can match sentences in an interpretable way.

One typical approach to long-form document matching is first conducting alignment between cross-document sentence pairs, and then aggregating all of the sentence-level matching signals. However, this approach could be problematic because the alignment between documents is partial — despite two documents as a whole are well-matched, most of the sentences could still be dissimilar. Those dissimilar sentences lead to spurious sentence-level matching signals which may overwhelm the real ones, increasing the difficulties of learning the matching function. Therefore, accurately selecting the key sentences for document matching is becoming a challenging issue. To address the issue, we propose a novel matching approach that equips existing document matching models with an Optimal Partial Transport (OPT) based component, namely OPT-Match, which selects the sentences that play a major role in matching. Enjoying the partial transport properties of OPT, the selected key sentences can not only effectively enhance the matching accuracy, but also be explained as the rationales for the matching results. Extensive experiments on four publicly available datasets demonstrated that existing methods equipped with OPT-Match consistently outperformed the corresponding underlying methods. Evaluations also showed that the key sentences selected by OPT-Match were consistent with human-provided rationales.

In this paper, we focus on the Audio-Visual Question Answering (AVQA) task, which aims to answer questions regarding different visual objects, sounds, and their associations in videos. The problem requires comprehensive multimodal understanding and spatio-temporal reasoning over audio-visual scenes. To benchmark this task and facilitate our study, we introduce a large-scale MUSIC-AVQA dataset, which contains more than 45K question-answer pairs covering 33 different question templates spanning over different modalities and question types. We develop several baselines and introduce a spatio-temporal grounded audio-visual network for the AVQA problem. Our results demonstrate that AVQA benefits from multisensory perception and our model outperforms recent A-, V-, and AVQA approaches. We believe that our built dataset has the potential to serve as testbed for evaluating and promoting progress in audio-visual scene understanding and spatio-temporal reasoning. Code and dataset: http://gewu-lab.github.io/MUSIC-AVQA/

Large-scale single-stream pre-training has shown dramatic performance in image-text retrieval. Regrettably, it faces low inference efficiency due to heavy attention layers. Recently, two-stream methods like CLIP and ALIGN with high inference efficiency have also shown promising performance, however, they only consider instance-level alignment between the two streams (thus there is still room for improvement). To overcome these limitations, we propose a novel COllaborative Two-Stream vision-language pretraining model termed COTS for image-text retrieval by enhancing cross-modal interaction. In addition to instance-level alignment via momentum contrastive learning, we leverage two extra levels of cross-modal interactions in our COTS: (1) Token-level interaction - a masked vision-language modeling (MVLM) learning objective is devised without using a cross-stream network module, where variational autoencoder is imposed on the visual encoder to generate visual tokens for each image. (2) Task-level interaction - a KL-alignment learning objective is devised between text-to-image and image-to-text retrieval tasks, where the probability distribution per task is computed with the negative queues in momentum contrastive learning. Under a fair comparison setting, our COTS achieves the highest performance among all two-stream methods and comparable performance (but with 10,800× faster in inference) w.r.t. the latest single-stream methods. Importantly, our COTS is also applicable to text-to-video retrieval, yielding new state-of-the-art on the widely-used MSR-VTT dataset.

We study the text generation task under the approach of pre-trained language models (PLMs).Typically, an auto-regressive (AR) method is adopted for generating texts in a token-bytoken manner.Despite many advantages of AR generation, it usually suffers from inefficient inference.Therefore, non-autoregressive (NAR) models are proposed to generate all target tokens simultaneously.However, NAR models usually generate texts of lower quality due to the absence of token dependency in the output text.In this paper, we propose ELMER: an Efficient and effective PLM for NAR tExt geneRation to explicitly model the token dependency during NAR generation.By leveraging the early exit technique, ELMER enables the token generations at different layers, according to their prediction confidence (a more confident token will exit at a lower layer).Besides, we propose a novel pre-training objective, Layer Permutation Language Modeling, to pre-train ELMER by permuting the exit layer for each token in sequences.Experiments on three text generation tasks show that ELMER significantly outperforms NAR models and further narrows the performance gap with AR PLMs (e.g., ELMER (29.92) vs BART (30.61)ROUGE-L in XSUM) while achieving over 10 times inference speedup.

To develop effective and efficient graph similarity learning (GSL) models, a series of data-driven neural algorithms have been proposed in recent years. Although GSL models are frequently deployed in privacy-sensitive scenarios, the user privacy protection of neural GSL models has not drawn much attention. To comprehensively understand the privacy protection issues, we first introduce the concept of attackable representation to systematically characterize the privacy attacks that each model can face. Inspired by the qualitative results, we propose a novel P rivacy-P reserving neural G raph M atching network model, named PPGM, for graph similarity learning. To prevent reconstruction attacks, the proposed model does not communicate node-level representations between devices. Instead, we learn multi-perspective graph representations based on learnable context vectors. To alleviate the attacks to graph properties, the obfuscated features that contain information from both graphs are communicated. In this way, the private properties of each graph can be difficult to infer. Based on the node-graph matching techniques while calculating the obfuscated features, PPGM can also be effective in similarity measuring. To quantitatively evaluate the privacy-preserving ability of neural GSL models, we further propose an evaluation protocol via training supervised black-box attack models. Extensive experiments on widely-used benchmarks show the effectiveness and strong privacy-protection ability of the proposed model PPGM. The code is available at: https://github.com/RUCAIBox/PPGM.

We propose a video feature representation learning frame-work called STAR-GNN, which applies a pluggable graph neural network component on a multi-scale lattice feature graph. The essence of STAR-GNN is to exploit both the temporal dynamics and spatial contents as well as vi-sual connections between regions at different scales in the frames. It models a video with a lattice feature graph in which the nodes represent regions of different granularity, with weighted edges that represent the spatial and temporal links. The contextual nodes are aggregated simultaneously by graph neural networks with parameters trained with re-trieval triplet loss. In the experiments, we show that STAR-GNN effectively implements a dynamic attention mechanism on video frame sequences, resulting in the emphasis for dy-namic and semantically rich content in the video, and is robust to noise and redundancies. Empirical results show that STAR-GNN achieves state-of-the-art performance for Content-Based Video Retrieval.

In order to develop effective sequential recommenders, a series of sequence representation learning (SRL) methods are proposed to model historical user behaviors. Most existing SRL methods rely on explicit item IDs for developing the sequence models to better capture user preference. Though effective to some extent, these methods are difficult to be transferred to new recommendation scenarios, due to the limitation by explicitly modeling item IDs. To tackle this issue, we present a novel universal sequence representation learning approach, named UniSRec. The proposed approach utilizes the associated description text of items to learn transferable representations across different recommendation scenarios. For learning universal item representations, we design a lightweight item encoding architecture based on parametric whitening and mixture-of-experts enhanced adaptor. For learning universal sequence representations, we introduce two contrastive pre-training tasks by sampling multi-domain negatives. With the pre-trained universal sequence representation model, our approach can be effectively transferred to new recommendation domains or platforms in a parameter-efficient way, under either inductive or transductive settings. Extensive experiments conducted on real-world datasets demonstrate the effectiveness of the proposed approach. Especially, our approach also leads to a performance improvement in a cross-platform setting, showing the strong transferability of the proposed universal SRL method. The code and pre-trained model are available at: https://github.com/RUCAIBox/UniSRec.

Relevant recommendation is a special recommendation scenario which provides relevant items when users express interests on one target item (e.g., click, like and purchase). Besides considering the relevance between recommendations and trigger item, the recommendations should also be diversified to avoid information cocoons. However, existing diversified recommendation methods mainly focus on item-level diversity which is insufficient when the recommended items are all relevant to the target item. Moreover, redundant or noisy item features might affect the performance of simple feature-aware recommendation approaches. Faced with these issues, we propose a Feature Disentanglement Self-Balancing Re-ranking framework (FDSB) to capture feature- aware diversity. The framework consists of two major modules, namely disentangled attention encoder (DAE) and self-balanced multi-aspect ranker. In DAE, we use multi-head attention to learn disentangled aspects from rich item features. In the ranker, we develop an aspect-specific ranking mechanism that is able to adaptively balance the relevance and diversity for each aspect. In experiments, we conduct offline evaluation on the collected dataset and deploy FDSB on KuaiShou app for online ??/?? test on the function of relevant recommendation. The significant improvements on both recommendation quality and user experience verify the effectiveness of our approach.

Search result diversification focuses on reducing redundancy and improving subtopic richness in the results for a given query. Most existing approaches measure document diversity mainly based on text or pre-trained representations. However, some underlying relationships between the query and documents are difficult for the model to capture only from the content. Given that the knowledge base can offer well-defined entities and explicit relationships between entities, we exploit knowledge to model the relationship between documents and the query and propose a knowledge-enhanced search result diversification approach KEDIV. Concretely, we build a query-specific relation graph to model the complicated query-document relationship from an entity view. Then a graph neural network and node weight adjust algorithm are applied to the relation graph to obtain context-aware entity representations and document representations at each selection step. The diversity features are derived from the updated node representations of the relation graph. In this way, we can take advantage of entities' abundant information to model document's diversity in search result diversification. Experimental results on commonly used datasets show that our proposed approach can outperform the state-of-the-art methods.

Conversational recommender systems (CRS) aim to proactively elicit user preference and recommend high-quality items through natural language conversations. Typically, a CRS consists of a recommendation module to predict preferred items for users and a conversation module to generate appropriate responses. To develop an effective CRS, it is essential to seamlessly integrate the two modules. Existing works either design semantic alignment strategies, or share knowledge resources and representations between the two modules. However, these approaches still rely on different architectures or techniques to develop the two modules, making it difficult for effective module integration. To address this problem, we propose a unified CRS model named UniCRS based on knowledge-enhanced prompt learning. Our approach unifies the recommendation and conversation subtasks into the prompt learning paradigm, and utilizes knowledge-enhanced prompts based on a fixed pre-trained language model (PLM) to fulfill both subtasks in a unified approach. In the prompt design, we include fused knowledge representations, task-specific soft tokens, and the dialogue context, which can provide sufficient contextual information to adapt the PLM for the CRS task. Besides, for the recommendation subtask, we also incorporate the generated response template as an important part of the prompt, to enhance the information interaction between the two subtasks. Extensive experiments on two public CRS datasets have demonstrated the effectiveness of our approach. Our code is publicly available at the link: https://github.com/RUCAIBox/UniCRS.

In streaming media applications, like music Apps, songs are recommended in a continuous way in users' daily life. The recommended songs are played automatically although users may not pay any attention to them, posing a challenge of user attention bias in training recommendation models, i.e., the training instances contain a large number of false-positive labels (users' feedback). Existing approaches either directly use the auto-feedbacks or heuristically delete the potential false-positive labels. Both of the approaches lead to biased results because the false-positive labels cause the shift of training data distribution, hurting the accuracy of the recommendation models. In this paper, we propose a learning-based counterfactual approach to adjusting the user auto-feedbacks and learning the recommendation models using Neural Dueling Bandit algorithm, called NDB. Specifically, NDB maintains two neural networks: a user attention network for computing the importance weights that are used for modifying the original rewards, and another random network trained with dueling bandit for conducting online recommendations based on the modified rewards. Theoretical analysis showed that the modified rewards are statistically unbiased, and the learned bandit policy enjoys a sub-linear regret bound. Experimental results demonstrated that NDB can significantly outperform the state-of-the-art baselines.

This paper aims to advance the mathematical intelligence of machines by presenting the first Chinese mathematical pre-trained language model (PLM) for effectively understanding and representing mathematical problems. Unlike other standard NLP tasks, mathematical texts are difficult to understand, since they involve mathematical terminology, symbols and formulas in the problem statement. Typically, it requires complex mathematical logic and background knowledge for solving mathematical problems.

Junyi Li, Tianyi Tang, Zheng Gong, Lixin Yang, Zhuohao Yu, Zhipeng Chen, Jingyuan Wang, Xin Zhao, Ji-Rong Wen. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

Junyi Li, Tianyi Tang, Jian-Yun Nie, Ji-Rong Wen, Xin Zhao. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

Hanxun Zhong, Zhicheng Dou, Yutao Zhu, Hongjin Qian, Ji-Rong Wen. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

Convolutional neural networks use regular quadrilateral convolution kernels to extract features. Since the number of parameters increases quadratically with the size of the convolution kernel, many popular models use small convolution kernels, resulting in small local receptive fields in lower layers. This paper proposes a novel log-polar space convolution (LPSC) layer, where the convolution kernel is elliptical and adaptively divides its local receptive field into different regions according to the relative directions and logarithmic distances. The local receptive field grows exponentially with the number of distance levels. Therefore, the proposed LPSC not only naturally encodes local spatial structures, but also greatly increases the single-layer receptive field while maintaining the number of parameters. We show that LPSC can be implemented with conventional convolution via log-polar space pooling and can be applied in any network architecture to replace conventional convolutions. Experiments on different tasks and datasets demonstrate the effectiveness of the proposed LPSC.

Designing spectral convolutional networks is a challenging problem in graph learning. ChebNet, one of the early attempts, approximates the spectral graph convolutions using Chebyshev polynomials. GCN simplifies ChebNet by utilizing only the first two Chebyshev polynomials while still outperforming it on real-world datasets. GPR-GNN and BernNet demonstrate that the Monomial and Bernstein bases also outperform the Chebyshev basis in terms of learning the spectral graph convolutions. Such conclusions are counter-intuitive in the field of approximation theory, where it is established that the Chebyshev polynomial achieves the optimum convergent rate for approximating a function. In this paper, we revisit the problem of approximating the spectral graph convolutions with Chebyshev polynomials. We show that ChebNet's inferior performance is primarily due to illegal coefficients learnt by ChebNet approximating analytic filter functions, which leads to over-fitting. We then propose ChebNetII, a new GNN model based on Chebyshev interpolation, which enhances the original Chebyshev polynomial approximation while reducing the Runge phenomenon. We conducted an extensive experimental study to demonstrate that ChebNetII can learn arbitrary graph convolutions and achieve superior performance in both full- and semi-supervised node classification tasks. Most notably, we scale ChebNetII to a billion graph ogbn-papers100M, showing that spectral-based GNNs have superior performance. Our code is available at https://github.com/ivam-he/ChebNetII.

AI-empowered drug recommendation has become an important task in healthcare research areas, which offers an additional perspective to assist human doctors with more accurate and more efficient drug prescriptions. Generally, drug recommendation is based on patients' diagnosis results in the electronic health records. We assume that there are three key factors to be addressed in drug recommendation: 1) elimination of recommendation bias due to limitations of observable information, 2) better utilization of historical health condition and 3) coordination of multiple drugs to control safety. To this end, we propose DrugRec, a causal inference based drug recommendation model. The causal graphical model can identify and deconfound the recommendation bias with front-door adjustment. Meanwhile, we model the multi-visit in the causal graph to characterize a patient's historical health conditions. Finally, we model the drug-drug interactions (DDIs) as the propositional satisfiability (SAT) problem, and solving the SAT problem can help better coordinate the recommendation. Comprehensive experiment results show that our proposed model achieves state-of-the-art performance on the widely used datasets MIMIC-III and MIMIC-IV, demonstrating the effectiveness and safety of our method.

Person-job fit is the core technique of online recruitment platforms, which can improve the efficiency of recruitment by accurately matching the job positions with the job seekers. Existing works mainly focus on modeling the unidirectional process or overall matching. However, recruitment is a two-way selection process, which means that both candidate and employer involved in the interaction should meet the expectation of each other, instead of unilateral satisfaction. In this paper, we propose a dual-perspective graph representation learning approach to model directed interactions between candidates and jobs. To model the two-way selection preference from the dual-perspective of job seekers and employers, we incorporate two different nodes for each candidate (or job) and characterize both successful matching and failed matching via a unified dual-perspective interaction graph. To learn dual-perspective node representations effectively, we design an effective optimization algorithm, which involves a quadruple-based loss and a dual-perspective contrastive learning loss. Extensive experiments on three large real-world recruitment datasets have shown the effectiveness of our approach. Our code is available at https://github.com/RUCAIBox/DPGNN .

Recently, graph neural networks (GNNs) have been shown powerful capacity at modeling structural data. However, when adapted to downstream tasks, it usually requires abundant task-specific labeled data, which can be extremely scarce in practice. A promising solution to data scarcity is to pre-train a transferable and expressive GNN model on large amounts of unlabeled graphs or coarse-grained labeled graphs. Then the pre-trained GNN is fine-tuned on downstream datasets with task-specific fine-grained labels. In this paper, we present a novel Graph Matching based GNN Pre-Training framework, called GMPT. Focusing on a pair of graphs, we propose to learn structural correspondences between them via neural graph matching, consisting of both intra-graph message passing and inter-graph message passing. In this way, we can learn adaptive representations for a given graph when paired with different graphs, and both node- and graph-level characteristics are naturally considered in a single pre-training task. The proposed method can be applied to fully self-supervised pre-training and coarse-grained supervised pre-training. We further propose an approximate contrastive training strategy to significantly reduce time/memory consumption. Extensive experiments on multi-domain, out-of-distribution benchmarks have demonstrated the effectiveness of our approach. The code is available at: https://github.com/RUCAIBox/GMPT.

Limited by the statistical-based machine learning framework, a spurious correlation is likely to appear in existing knowledge-aware recommendation methods. It refers to a knowledge fact that appears causal to the user behaviors (inferred by the recommender) but is not in fact. For tackling this issue, we present a novel approach to discovering and alleviating the potential spurious correlations from a counterfactual perspective. To be specific, our approach consists of two counterfactual generators and a recommender. The counterfactual generators are designed to generate counterfactual interactions via reinforcement learning, while the recommender is implemented with two different graph neural networks to aggregate the information from KG and user-item interactions respectively. The counterfactual generators and recommender are integrated in a mutually collaborative way. With this approach, the recommender helps the counterfactual generators better identify potential spurious correlations and generate high-quality counterfactual interactions, while the counterfactual generators help the recommender weaken the influence of the potential spurious correlations simultaneously. Extensive experiments on three real-world datasets have shown the effectiveness of the proposed approach by comparing it with a number of competitive baselines. Our implementation code is available at: https://github.com/RUCAIBox/CGKR.

As an essential operation of legal retrieval, legal case matching plays a central role in intelligent legal systems. This task has a high demand on the explainability of matching results because of its critical impacts on downstream applications --- the matched legal cases may provide supportive evidence for the judgments of target cases and thus influence the fairness and justice of legal decisions. Focusing on this challenging task, we propose a novel and explainable method, namely IOT-Match, with the help of computational optimal transport, which formulates the legal case matching problem as an inverse optimal transport (IOT) problem. Different from most existing methods, which merely focus on the sentence-level semantic similarity between legal cases, our IOT-Match learns to extract rationales from paired legal cases based on both semantics and legal characteristics of their sentences. The extracted rationales are further applied to generate faithful explanations and conduct matching. Moreover, the proposed IOT-Match is robust to the alignment label insufficiency issue commonly in practical legal case matching tasks, which is suitable for both supervised and semi-supervised learning paradigms. To demonstrate the superiority of our IOT-Match method and construct a benchmark of explainable legal case matching task, we not only extend the well-known Challenge of AI in Law (CAIL) dataset but also build a new Explainable Legal cAse Matching (ELAM) dataset, which contains lots of legal cases with detailed and explainable annotations. Experiments on these two datasets show that our IOT-Match outperforms state-of-the-art methods consistently on matching prediction, rationale extraction, and explanation generation.

Asking clarifying questions is an interactive way to effectively clarify user intent. When a user submits a query, the search engine will return a clarifying question with several clickable items of sub-intents for clarification. According to the existing definition, the key to asking high-quality questions is to generate good descriptions for submitted queries and provided items. However, existing methods mainly based on static knowledge bases are difficult to find descriptions for many queries because of the lack of entities within these queries and their corresponding items. For such a query, it is unable to generate an informative question. To alleviate this problem, we propose leveraging top search results of the query to help generate better descriptions because we deem that the top retrieved documents contain rich and relevant contexts of the query. Specifically, we first design a rule-based algorithm to extract description candidates from search results and rank them by various human-designed features. Then, we apply an learning-to-rank model and another generative model for generalization and further improve the quality of clarifying questions. Experimental results show that our proposed methods can generate more readable and informative questions compared with existing methods. The results prove that search results can be utilized to improve users' search experience for search clarification in conversational search systems.

In session search, it's important to utilize historical interactions between users and the search engines to improve document retrieval. However, not all historical information contributes to document ranking. Users often express their preferences in the process of modifying the previous query, which can help us catch useful information in the historical interactions. Inspired by it, we propose to model historical query change to improve document ranking performance. Especially, we characterize multi-granularity query change between each pair of adjacent queries at both term level and semantic level. For term level query change, we calculate three types of term weights, including the retained term weights, added term weights and removed term weights. Then we perform term-based interaction between the candidate document and historical queries based on the term weights. For semantic level query change, we calculate an overall representation of user intent by integrating the representations of each historical query obtained by different types of term weights. Then we adopt representation-based matching between this representation and the candidate document. To improve the effect of query change modeling, we introduce query change classification as an auxiliary task. Experimental results on AOL and TianGong-ST search logs show that our model outperforms most existing models for session search.

Product search has been an important way for people to find products on online shopping platforms. Existing approaches in personalized product search mainly embed user preferences into one single vector. However, this simple strategy easily results in sub-optimal representations, failing to model and disentangle user’s multiple preferences. To overcome this problem, we proposed a category-aware multi-interest model to encode users as multiple preference embeddings to represent user-specific interests. Specifically, we also capture the category indications for each preference to indicate the distribution of categories it focuses on, which is derived from rich relations between users, products, and attributes. Based on these category indications, we develop a category attention mechanism to aggregate these various preference embeddings considering current queries and items as the user’s comprehensive representation. By this means, we can use this representation to calculate matching scores of retrieved items to determine whether they meet the user’s search intent. Besides, we introduce a homogenization regularization term to avoid the redundancy between user interests. Experimental results show that the proposed method significantly outperforms existing approaches.

Ranking has been one of the most important tasks in information retrieval. With the development of deep representation learning, many researchers propose to encode both the query and items into embedding vectors and rank the items according to the inner product or distance measures in the embedding space. However, the ranking models based on vector embeddings may have shortages in effectiveness and efficiency. For effectiveness, they lack the intrinsic ability to model the diversity and uncertainty of queries and items in ranking. For efficiency, nearest neighbor search in a large collection of item vectors can be costly. In this work, we propose to use the recently proposed probabilistic box embeddings for effective and efficient ranking, in which queries and items are parameterized as high-dimensional axis-aligned hyper-rectangles. For effectiveness, we utilize probabilistic box embeddings to model the diversity and uncertainty with the overlapping relations of the hyper-rectangles, and prove that such overlapping measure is a kernel function which can be adopted in other kernel-based methods. For efficiency, we propose a box embedding-based indexing method, which can safely filter irrelevant items and reduce the retrieval latency. We further design a training strategy to increase the proportion of irrelevant items that can be filtered by the index. Experiments on public datasets show that the box embeddings and the box embedding-based indexing approaches are effective and efficient in two ranking tasks: ad hoc retrieval and product recommendation.

Machine-learning based recommender system(RS) has become an effective means to help people automatically discover their interests. Existing models often represent the rich information for recommendation, such as items, users, and contexts, as embedding vectors and leverage them to predict users’ feedback. In the view of causal analysis, the associations between these embedding vectors and users’ feedback are a mixture of the causal part that describes why an item is preferred by a user, and the non-causal part that merely reflects the statistical dependencies between users and items, for example, the exposure mechanism, public opinions, display position, etc. However, existing RSs mostly ignored the striking differences between the causal parts and non-causal parts when using these embedding vectors. In this paper, we propose a model-agnostic framework named IV4Rec that can effectively decompose the embedding vectors into these two parts, hence enhancing recommendation results. Specifically, we jointly consider users’ behaviors in search scenarios and recommendation scenarios. Adopting the concepts in causal analysis, we embed users’ search behaviors as instrumental variables (IVs), to help decompose original embedding vectors in recommendation, i.e., treatments. IV4Rec then combines the two parts through deep neural networks and uses the combined results for recommendation. IV4Rec is model-agnostic and can be applied to a number of existing RSs such as DIN and NRHUB. Experimental results on both public and proprietary industrial datasets demonstrate that IV4Rec consistently enhances RSs and outperforms a framework that jointly considers search and recommendation.

Sequential recommendation holds the promise of understanding user preference by capturing successive behavior correlations. Existing research focus on designing different models for better fitting the offline datasets. However, the observational data may have been contaminated by the exposure or selection biases, which renders the learned sequential models unreliable. In order to solve this fundamental problem, in this paper, we propose to reformulate the sequential recommendation task with the potential outcome framework, where we are able to clearly understand the data bias mechanism and correct it by re-weighting the training instances with the inverse propensity score (IPS). For more robustness modeling, a clipping strategy is applied to the IPS estimation to reduce the variance of the learning objective. To make our framework more practical, we design a parameterized model to remove the impact of the potential latent confounders. At last, we theoretically analyze the unbiasedness of the proposed framework under both vanilla and clipping IPS estimations. To the best of our knowledge, this is the first work on debiased sequential recommendation. We conduct extensive experiment based on both synthetic and real-world datasets to demonstrate the effectiveness of our framework.

Recently, deep neural networks such as RNN, CNN and Transformer have been applied in the task of sequential recommendation, which aims to capture the dynamic preference characteristics from logged user behavior data for accurate recommendation. However, in online platforms, logged user behavior data is inevitable to contain noise, and deep recommendation models are easy to overfit on these logged data. To tackle this problem, we borrow the idea of filtering algorithms from signal processing that attenuates the noise in the frequency domain. In our empirical experiments, we find that filtering algorithms can substantially improve representative sequential recommendation models, and integrating simple filtering algorithms (e.g., Band-Stop Filter) with an all-MLP architecture can even outperform competitive Transformer-based models. Motivated by it, we propose FMLP-Rec, an all-MLP model with learnable filters for sequential recommendation task. The all-MLP architecture endows our model with lower time complexity, and the learnable filters can adaptively attenuate the noise information in the frequency domain. Extensive experiments conducted on eight real-world datasets demonstrate the superiority of our proposed method over competitive RNN, CNN, GNN and Transformer-based methods. Our code and data are publicly available at the link: https://github.com/RUCAIBox/FMLP-Rec .

Audiovisual scenes are pervasive in our daily life. It is commonplace for humans to discriminatively localize different sounding objects but quite challenging for machines to achieve class-aware sounding objects localization without category annotations, i.e., localizing the sounding object and recognizing its category. To address this problem, we propose a two-stage step-by-step learning framework to localize and recognize sounding objects in complex audiovisual scenarios using only the correspondence between audio and vision. First, we propose to determine the sounding area via coarse-grained audiovisual correspondence in the single source cases. Then visual features in the sounding area are leveraged as candidate object representations to establish a category-representation object dictionary for expressive visual character extraction. We generate class-aware object localization maps in cocktail-party scenarios and use audiovisual correspondence to suppress silent areas by referring to this dictionary. Finally, we employ category-level audiovisual consistency as the supervision to achieve fine-grained audio and sounding object distribution alignment. Experiments on both realistic and synthesized videos show that our model is superior in localizing and recognizing objects as well as filtering out silent ones. We also transfer the learned audiovisual network into the unsupervised object detection task, obtaining reasonable performance.

Supervised dimensionality reduction for sequence data learns a transformation that maps the observations in sequences onto a low-dimensional subspace by maximizing the separability of sequences in different classes. It is typically more challenging than conventional dimensionality reduction for static data, because measuring the separability of sequences involves non-linear procedures to manipulate the temporal structures. In this paper, we propose a linear method, called order-preserving Wasserstein discriminant analysis (OWDA), and its deep extension, namely DeepOWDA, to learn linear and non-linear discriminative subspace for sequence data, respectively. We construct novel separability measures between sequence classes based on the order-preserving Wasserstein (OPW) distance to capture the essential differences among their temporal structures. Specifically, for each class, we extract the OPW barycenter and construct the intra-class scatter as the dispersion of the training sequences around the barycenter. The inter-class distance is measured as the OPW distance between the corresponding barycenters. We learn the linear and non-linear transformations by maximizing the inter-class distance and minimizing the intra-class scatter. In this way, the proposed OWDA and DeepOWDA are able to concentrate on the distinctive differences among classes by lifting the geometric relations with temporal constraints. Experiments on four 3D action recognition datasets show the effectiveness of OWDA and DeepOWDA.

Personalized PageRank (PPR) is a popular node proximity metric in graph mining and network research. A single-source PPR (SSPPR) query asks for the PPR value of each node on the graph. Due to its importance and wide applications, decades of efforts have been devoted to the efficient processing of SSPPR queries. Among existing algorithms, LocalPush is a fundamental method for SSPPR queries and serves as a cornerstone for subsequent algorithms. In LocalPush , a push operation is a crucial primitive operation, which distributes the probability at a node u to ALL u 's neighbors via the corresponding edges. Although this push operation works well on unweighted graphs, unfortunately, it can be rather inefficient on weighted graphs. In particular, on unbalanced weighted graphs where only a few of these edges take the majority of the total weight among them, the push operation would have to distribute "insignificant" probabilities along those edges which just take the minor weights, resulting in expensive overhead. To resolve this issue, in this paper, we propose the EdgePush algorithm, a novel method for computing SSPPR queries on weighted graphs. EdgePush decomposes the aforementioned push operations in edge-based push , allowing the algorithm to operate at the edge level granularity. As a result, it can flexibly distribute the probabilities according to edge weights. Furthermore, our EdgePush allows a fine-grained termination threshold for each individual edge, leading to a superior complexity over LocalPush. Notably, we prove that EdgePush improves the theoretical query cost of LocalPush by an order of up to O ( n ) when the graph's weights are unbalanced. Our experimental results demonstrate that EdgePush significantly outperforms state-of-the-art baselines in terms of query efficiency on large motif-based and real-world weighted graphs.

Projecting the input text pair into a common semantic space where the matching function can be readily learned is an essential step for asymmetrical text matching. In the practice, it is often observed that the feature vectors from asymmetrical texts show a tendency to be gradually undistinguishable in the semantic space as the model is trained. However, the phenomenon is overlooked in existing studies. As a result, the feature vectors are constructed without any regularization, which inevitably hinders the learning of the downstream matching functions. In this paper, we first exploit the phenomenon and propose DDR-Match, a novel matching framework tailored for asymmetrical text matching. Specifically, in DDR-Match, a distribution distance-based regularizer is devised to accelerate the fusion of sequence representations corresponding to different domains in the semantic space. Then, we provide three instances of DDR-Match and make a comparison among them. DDR-Match is compatible with existing text matching methods by incorporating them as the underlying matching model. Four popular text matching methods are exploited in the paper. Extensive experimental results based on five publicly available benchmarks showed that DDR-Match consistently outperformed its underlying methods.

Active learning is an effective approach for tasks with limited labeled data. It samples a small set of data to annotate actively and is widely applied in various AI tasks. It uses an iterative process, during which we utilize the current trained model to evaluate all unlabeled samples and annotate the best samples based on a specific query strategy to update the underlying model iteratively. Most existing active learning approaches rely on only the evaluation results generated by the current model and ignore the results from previous iterations. In this paper, we propose using more historical evaluation results which can provide additional information to help better select samples. First, we apply two kinds of heuristic features of the historical evaluation results, the weighted sum of historical results and the fluctuation of the historical evaluation sequence, to improve the effectiveness of active learning sampling. Next, to further and more globally use the information contained in the historical results, we design a novel query strategy that learns how to select samples based on the historical sequences automatically. Our proposed idea is general and can be combined with both basic and state-of-the-art query strategies to achieve improvements. We test our approaches on two common NLP tasks including text classification and named entity recognition. Experimental results show that our methods significantly promote existing methods.

A persistent data structure , also known as a multiversion data structure in the database literature, is a data structure that preserves all its previous versions as it is updated over time. Every update (inserting, deleting, or changing a data record) to the data structure creates a new version, while all the versions are kept in the data structure so that any previous version can still be queried. Persistent data structures aim at recording all versions accurately, which results in a space requirement that is at least linear to the number of updates. In many of today’s big data applications, in particular, for high-speed streaming data, the volume and velocity of the data are so high that we cannot afford to store everything. Therefore, streaming algorithms have received a lot of attention in the research community, which uses only sublinear space by sacrificing slightly on accuracy. All streaming algorithms work by maintaining a small data structure in memory, which is usually called a sketch , summary , or synopsis . The summary is updated upon the arrival of every element in the stream, thus it is ephemeral , meaning that it can only answer queries about the current status of the stream. In this article, we aim at designing persistent summaries, thereby giving streaming algorithms the ability to answer queries about the stream at any prior time.

Citation count prediction is an important task for estimating the future impact of research papers. Most of the existing works utilize the information extracted from the paper itself. In this article, we focus on how to utilize another kind of useful data signal (i.e., peer review text) to improve both the performance and interpretability of the prediction models. Specially, we propose a novel aspect-aware capsule network for citation count prediction based on review text. It contains two major capsule layers, namely the feature capsule layer and the aspect capsule layer, with two different routing approaches, respectively. Feature capsules encode the local semantics from review sentences as the input of aspect capsule layer, whereas aspect capsules aim to capture high-level semantic features that will be served as final representations for prediction. Besides the predictive capacity, we also enhance the model interpretability with two strategies. First, we use the topic distribution of the review text to guide the learning of aspect capsules so that each aspect capsule can represent a specific aspect in the review. Then, we use the learned aspect capsules to generate readable text for explaining the predicted citation count. Extensive experiments on two real-world datasets have demonstrated the effectiveness of the proposed model in both performance and interpretability.

In recommender systems, it is essential to understand the underlying factors that affect user-item interaction. Recently, several studies have utilized disentangled representation learning to discover such hidden factors from user-item interaction data, which shows promising results. However, without any external guidance signal, the learned disentangled representations lack clear meanings, and are easy to suffer from the data sparsity issue. In light of these challenges, we study how to leverage knowledge graph (KG) to guide the disentangled representation learning in recommender systems. The purpose for incorporating KG is twofold, making the disentangled representations interpretable and resolving data sparsity issue. However, it is not straightforward to incorporate KG for improving disentangled representations, because KG has very different data characteristics compared with user-item interactions. We propose a novel K nowledge-guided D isentangled R epresentations approach ( KDR ) to utilizing KG to guide the disentangled representation learning in recommender systems. The basic idea, is to first learn more interpretable disentangled dimensions (explicit disentangled representations) based on structural KG, and then align implicit disentangled representations learned from user-item interaction with the explicit disentangled representations. We design a novel alignment strategy based on mutual information maximization. It enables the KG information to guide the implicit disentangled representation learning, and such learned disentangled representations will correspond to semantic information derived from KG. Finally, the fused disentangled representations are optimized to improve the recommendation performance. Extensive experiments on three real-world datasets demonstrate the effectiveness of the proposed model in terms of both performance and interpretability.

Personalized search tailors document ranking lists for each individual user based on her interests and query intent to better satisfy the user’s information need. Many personalized search models have been proposed. They first build a user interest profile from the user’s search history, and then re-rank the documents based on the personalized matching scores between the created profile and candidate documents. In this article, we attempt to solve the personalized search problem from an alternative perspective of clarifying the user’s intention of the current query. We know that there are many ambiguous words in natural language such as “Apple.” People with different knowledge backgrounds and interests have personalized understandings of these words. Therefore, we propose a personalized search model with personal word embeddings for each individual user that mainly contain the word meanings that the user already knows and can reflect the user interests. To learn great personal word embeddings, we design a pre-training model that captures both the textual information of the query log and the information about user interests contained in the click-through data represented as a graph structure. With personal word embeddings, we obtain the personalized word and context-aware representations of the query and documents. Furthermore, we also employ the current session as the short-term search context to dynamically disambiguate the current query. Finally, we use a matching model to calculate the matching score between the personalized query and document representations for ranking. Experimental results on two large-scale query logs show that our designed model significantly outperforms state-of-the-art personalization models.

Regret bounds of online kernel selection in a finite kernel set have been well studied, having at least an order O( √ NT) of magnitude after T rounds, where N is the number of candidate kernels. But it is still an unsolved problem to achieve sublinear regret bounds of online kernel selection in a continuous kernel space under different learning frameworks. In this paper, to represent different learning frameworks of online kernel selection, we divide online kernel selection approaches in a continuous kernel space into two categories according to the order of selection and training at each round. Then we construct a surrogate hypothesis space that contains all the candidate kernels with bounded norms and inner products, representing the continuously varying hypothesis space. Finally, we decompose the regrets of the proposed online kernel selection categories into different types of instantaneous regrets in the surrogate hypothesis space, and derive optimal regret bounds of order O( √ T) of magnitude under mild assumptions, independent of the cardinality of the continuous kernel space. Empirical studies verified the correctness of the theoretical regret analyses.

Peiyu Liu, Ze-Feng Gao, Wayne Xin Zhao, Zhi-Yuan Xie, Zhong-Yi Lu, Ji-Rong Wen. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Chongyang Tao, Changyu Chen, Jiazhan Feng, Ji-Rong Wen, Rui Yan. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Jingxuan Yang, Kerui Xu, Jun Xu, Si Li, Sheng Gao, Jun Guo, Nianwen Xue, Ji-Rong Wen. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Within online platforms, it is critical to capture the semantics of sequential user behaviors for accurately modeling user interests. However, dynamic characteristics and sparse behaviors make it difficult to train effective user representations for sequential user behavior modeling.

Designing pre-training objectives that more closely resemble the downstream tasks for pre-trained language models can lead to better performance at the fine-tuning stage, especially in the ad-hoc retrieval area. Existing pre-training approaches tailored for IR tried to incorporate weak supervised signals, such as query-likelihood based sampling, to construct pseudo query-document pairs from the raw textual corpus. However, these signals rely heavily on the sampling method. For example, the query likelihood model may lead to much noise in the constructed pre-training data. In this paper, we propose to leverage the large-scale hyperlinks and anchor texts to pre-train the language model for ad-hoc retrieval. Since the anchor texts are created by webmasters and can usually summarize the target document, it can help to build more accurate and reliable pre-training samples than a specific algorithm. Considering different views of the downstream ad-hoc retrieval, we devise four pre-training tasks based on the hyperlinks. We then pre-train the Transformer model to predict the pair-wise preference, jointly with the Masked Language Model objective. Experimental results on two large-scale ad-hoc retrieval datasets show the significant improvement of our model compared with the existing methods.

In this paper, we explore the problem of developing personalized chatbots. A personalized chatbot is designed as a digital chatting assistant for a user. The key characteristic of a personalized chatbot is that it should have a consistent personality with the corresponding user. It can talk the same way as the user when it is delegated to respond to others' messages. Many methods have been proposed to assign a personality to dialogue chatbots, but most of them utilize explicit user profiles, including several persona descriptions or key-value-based personal information. In a practical scenario, however, users might be reluctant to write detailed persona descriptions, and obtaining a large number of explicit user profiles requires tremendous manual labour. To tackle the problem, we present a retrieval-based personalized chatbot model, namely IMPChat, to learn an implicit user profile from the user's dialogue history. We argue that the implicit user profile is superior to the explicit user profile regarding accessibility and flexibility. IMPChat aims to learn an implicit user profile through modeling user's personalized language style and personalized preferences separately. To learn a user's personalized language style, we elaborately build language models from shallow to deep using the user's historical responses; To model a user's personalized preferences, we explore the conditional relations underneath each post-response pair of the user. The personalized preferences are dynamic and context-aware: we assign higher weights to those historical pairs that are topically related to the current query when aggregating the personalized preferences. We match each response candidate with the personalized language style and personalized preference, respectively, and fuse the two matching signals to determine the final ranking score. We conduct comprehensive experiments on two large datasets, and the results show that our method outperforms all baseline models.

Search and recommendation are the two most common approaches used by people to obtain information. They share the same goal -- satisfying the user's information need at the right time. There are already a lot of Internet platforms and Apps providing both search and recommendation services, showing us the demand and opportunity to simultaneously handle both tasks. However, most platforms consider these two tasks independently -- they tend to train separate search model and recommendation model, without exploiting the relatedness and dependency between them. In this paper, we argue that jointly modeling these two tasks will benefit both of them and finally improve overall user satisfaction. We investigate the interactions between these two tasks in the specific information content service domain. We propose first integrating the user's behaviors in search and recommendation into a heterogeneous behavior sequence, then utilizing a joint model for handling both tasks based on the unified sequence. More specifically, we design the Unified Information SEarch and Recommendation model (USER), which mines user interests from the integrated sequence and accomplish the two tasks in a unified way. Experiments on a dataset from a real-world information content service platform verify that our model outperforms separate search and recommendation baselines.

In recent years, there are a large number of recommendation algorithms proposed in the literature, from traditional collaborative filtering to deep learning algorithms. However, the concerns about how to standardize open source implementation of recommendation algorithms continually increase in the research community. In the light of this challenge, we propose a unified, comprehensive and efficient recommender system library called RecBole (pronounced as [rEk'[email protected]]), which provides a unified framework to develop and reproduce recommendation algorithms for research purpose. In this library, we implement 73 recommendation models on 28 benchmark datasets, covering the categories of general recommendation, sequential recommendation, context-aware recommendation and knowledge-based recommendation. We implement the RecBole library based on PyTorch, which is one of the most popular deep learning frameworks. Our library is featured in many aspects, including general and extensible data structures, comprehensive benchmark models and datasets, efficient GPU-accelerated execution, and extensive and standard evaluation protocols. We provide a series of auxiliary functions, tools, and scripts to facilitate the use of this library, such as automatic parameter tuning and break-point resume. Such a framework is useful to standardize the implementation and evaluation of recommender systems. The project and documents are released at https://recbole.io/.

Personalized search plays a crucial role in improving user search experience owing to its ability to build user profiles based on historical behaviors. Previous studies have made great progress in extracting personal signals from the query log and learning user representations. However, neural personalized search is extremely dependent on sufficient data to train the user model. Data sparsity is an inevitable challenge for existing methods to learn high-quality user representations. Moreover, the overemphasis on final ranking quality leads to rough data representations and impairs the generalizability of the model. To tackle these issues, we propose a Personalized Search framework with Self-supervised Learning (PSSL) to enhance data representations. Specifically, we adopt a contrastive sampling method to extract paired self-supervised information from sequences of user behaviors in query logs. Four auxiliary tasks are designed to pre-train the sentence encoder and the sequence encoder used in the ranking model. They are optimized by contrastive loss which aims to close the distance between similar user sequences, queries, and documents. Experimental results on two datasets demonstrate that our proposed model PSSL achieves state-of-the-art performance compared with existing baselines.

VQA models may tend to rely on language bias as a shortcut and thus fail to sufficiently learn the multi-modal knowledge from both vision and language. Recent debiasing methods proposed to exclude the language prior during inference. However, they fail to disentangle the "good" language context and "bad" language bias from the whole. In this paper, we investigate how to mitigate language bias in VQA. Motivated by causal effects, we proposed a novel counterfactual inference framework, which enables us to capture the language bias as the direct causal effect of questions on answers and reduce the language bias by subtracting the direct language effect from the total causal effect. Experiments demonstrate that our proposed counterfactual inference framework 1) is general to various VQA backbones and fusion strategies, 2) achieves competitive performance on the language-bias sensitive VQA-CP dataset while performs robustly on the balanced VQA v2 dataset without any augmented data. The code is available at https://github.com/yuleiniu/cfvqa.

In various natural language processing tasks, passage retrieval and passage re-ranking are two key procedures in finding and ranking relevant information.Since both the two procedures contribute to the final performance, it is important to jointly optimize them in order to achieve mutual improvement.In this paper, we propose a novel joint training approach for dense passage retrieval and passage reranking.A major contribution is that we introduce the dynamic listwise distillation, where we design a unified listwise training approach for both the retriever and the re-ranker.During the dynamic distillation, the retriever and the re-ranker can be adaptively improved according to each other's relevance information.We also propose a hybrid data augmentation strategy to construct diverse training instances for listwise training approach.Extensive experiments show the effectiveness of our approach on both MSMARCO and Natural Questions datasets.Our code is available at https:// github.com/PaddlePaddle/RocketQA.

Recent works have shown that powerful pretrained language models (PLM) can be fooled by small perturbations or intentional attacks.To solve this issue, various data augmentation techniques are proposed to improve the robustness of PLMs.However, it is still challenging to augment semantically relevant examples with sufficient diversity.In this work, we present Virtual Data Augmentation (VDA), a general framework for robustly fine-tuning PLMs.Based on the original token embeddings, we construct a multinomial mixture for augmenting virtual data embeddings, where a masked language model guarantees the semantic relevance and the Gaussian noise provides the augmentation diversity.Furthermore, a regularized training strategy is proposed to balance the two aspects.Extensive experiments on six datasets show that our approach is able to improve the robustness of PLMs and alleviate the performance degradation under adversarial attacks.

Complex action segmentation aims to detect what actions and when they happen in fine-grained level from long videos. Despite the fact that videos are often stored in a compressed format (e.g., MPEG-4), most existing approaches are proposed to directly model raw RGB videos: when only compressed videos are accessible, they have to first decode these videos, which is very time-consuming. In this paper, by explicitly leveraging the ‘compressed’ characteristic of compressed videos, we are the first to address the challenging task of complex action segmentation in compressed videos. To extract meaningful representations for complex action segmentation, we introduce the GOP-Level Compressed features (Golec), which can be obtained directly from compressed videos without video decompression. Importantly, by taking GOPs as the atomic units of actions, our Golec representation is intrinsically suitable for fine-grained action segmentation. Moreover, to remedy the coarser motion vectors (compared with optical flows which are computed from raw frames) used in our Golec representation for capturing the temporal context, we propose a new Bi-path knowledge distillation strategy. Extensive experiments show the effectiveness of our Golec representation and the Bi-path strategy. Importantly, our proposed model for complex action detection not only runs 5.2 times faster but also achieves significantly better results than the state-of-the-art alternatives using raw videos.

Knowledge base question answering (KBQA) aims to answer a question over a knowledge base (KB). Recently, a large number of studies focus on semantically or syntactically complicated questions. In this paper, we elaborately summarize the typical challenges and solutions for complex KBQA. We begin with introducing the background about the KBQA task. Next, we present the two mainstream categories of methods for complex KBQA, namely semantic parsing-based (SP-based) methods and information retrieval-based (IR-based) methods. We then review the advanced methods comprehensively from the perspective of the two categories. Specifically, we explicate their solutions to the typical challenges. Finally, we conclude and discuss some promising directions for future research.

Text generation has become one of the most important yet challenging tasks in natural language processing (NLP). The resurgence of deep learning has greatly advanced this field by neural generation models, especially the paradigm of pretrained language models (PLMs). In this paper, we present an overview of the major advances achieved in the topic of PLMs for text generation. As the preliminaries, we present the general task definition and briefly describe the mainstream architectures of PLMs for text generation. As the core content, we discuss how to adapt existing PLMs to model different input data and satisfy special properties in the generated text. We further summarize several important fine-tuning strategies for text generation. Finally, we present several future directions and conclude this paper. Our survey aims to provide text generation researchers a synthesis and pointer to related research.

As it becomes prevalent that user information exists in multiple platforms or services, cross-domain recommendation has been an important task in industry. Although it is well known that users tend to show different preferences in different domains, existing studies seldom model how domain biases affect user preferences. Focused on this issue, we develop a casual-based approach to mitigating the domain biases when transferring the user information cross domains. To be specific, this paper presents a novel debiasing learning based cross-domain recommendation framework with causal embedding. In this framework, we design a novel Inverse-Propensity-Score (IPS) estimator designed for cross-domain scenario, and further propose three kinds of restrictions for propensity score learning. Our framework can be generally applied to various recommendation algorithms for cross-domain recommendation. Extensive experiments on both public and industry datasets have demonstrated the effectiveness of the proposed framework.

Efficient computation of node proximity queries such as transition probabilities, Personalized PageRank, and Katz are of fundamental importance in various graph mining and learning tasks. In particular, several recent works leverage fast node proximity computation to improve the scalability of Graph Neural Networks (GNN). However, prior studies on proximity computation and GNN feature propagation are on a case-by-case basis, with each paper focusing on a particular proximity measure.

This work concerns self-supervised video representation learning (SSVRL), one topic that has received much attention recently. Since videos are storage-intensive and contain a rich source of visual content, models designed for SSVRL are expected to be storage- and computation-efficient, as well as effective. However, most existing methods only focus on one of the two objectives, failing to consider both at the same time. In this work, for the first time, the seemingly contradictory goals are simultaneously achieved by exploiting compressed videos and capturing mutual information between two input streams. Specifically, a novel Motion Vector based Cross Guidance Contrastive learning approach (MVCGC) is proposed. For storage and computation efficiency, we choose to directly decode RGB frames and motion vectors (that resemble low-resolution optical flows) from compressed videos on-the-fly. To enhance the representation ability of the motion vectors, hence the effectiveness of our method, we design a cross guidance contrastive learning algorithm based on multi-instance InfoNCE loss, where motion vectors can take supervision signals from RGB frames and vice versa. Comprehensive experiments on two downstream tasks show that our MVCGC yields new state-of-the-art while being significantly more efficient than its competitors.

Self-attention networks (SANs) have been intensively applied for sequential recommenders, but they are limited due to: (1) the quadratic complexity and vulnerability to over-parameterization in self-attention; (2) inaccurate modeling of sequential relations between items due to the implicit position encoding. In this work, we propose the low-rank decomposed self-attention networks (LightSANs) to overcome these problems. Particularly, we introduce the low-rank decomposed self-attention, which projects user's historical items into a small constant number of latent interests and leverages item-to-interest interaction to generate the context-aware representation. It scales linearly w.r.t. the user's historical sequence length in terms of time and space, and is more resilient to over-parameterization. Besides, we design the decoupled position encoding, which models the sequential relations between items more precisely. Extensive experimental studies are carried out on three real-world datasets, where LightSANs outperform the existing SANs-based recommenders in terms of both effectiveness and efficiency.

As a natural language generation task, it is challenging to generate informative and coherent review text. In order to enhance the informativeness of the generated text, existing solutions typically learn to copy entities or triples from knowledge graphs (KGs). However, they lack overall consideration to select and arrange the incorporated knowledge, which tends to cause text incoherence. To address the above issue, we focus on improving entity-centric coherence of the generated reviews by leveraging the semantic structure of KGs. In this paper, we propose a novel Coherence Enhanced Text Planning model (CETP) based on knowledge graphs (KGs) to improve both global and local coherence for review generation. The proposed model learns a two-level text plan for generating a document: (1) the document plan is modeled as a sequence of sentence plans in order, and (2) the sentence plan is modeled as an entity-based subgraph from KG. Local coherence can be naturally enforced by KG subgraphs through intra-sentence correlations between entities. For global coherence, we design a hierarchical self-attentive architecture with both subgraph- and node-level attention to enhance the correlations between subgraphs. To our knowledge, we are the first to utilize a KG-based text planning model to enhance text coherence for review generation. Extensive experiments on three datasets confirm the effectiveness of our model on improving the content coherence of generated texts.

Personalized chatbots focus on endowing chatbots with a consistent personality to behave like real users, give more informative responses, and further act as personal assistants. Existing personalized approaches tried to incorporate several text descriptions as explicit user profiles. However, the acquisition of such explicit profiles is expensive and time-consuming, thus being impractical for large-scale real-world applications. Moreover, the restricted predefined profile neglects the language behavior of a real user and cannot be automatically updated together with the change of user interests. In this paper, we propose to learn implicit user profiles automatically from large-scale user dialogue history for building personalized chatbots. Specifically, leveraging the benefits of Transformer on language understanding, we train a personalized language model to construct a general user profile from the user's historical responses. To highlight the relevant historical responses to the input post, we further establish a key-value memory network of historical post-response pairs, and build a dynamic post-aware user profile. The dynamic profile mainly describes what and how the user has responded to similar posts in history. To explicitly utilize users' frequently used words, we design a personalized decoder to fuse two decoding strategies, including generating a word from the generic vocabulary and copying one word from the user's personalized vocabulary. Experiments on two real-world datasets show the significant improvement of our model compared with existing methods.

atural language dialogue systems raise great attention recently. As many dialogue models are data-driven, high-quality datasets are essential to these systems. In this paper, we introduce Pchatbot, a large-scale dialogue dataset that contains two subsets collected from Weibo and Judicial forums respectively. To adapt the raw dataset to dialogue systems, we elaborately normalize the raw dataset via processes such as anonymization, deduplication, segmentation, and filtering. The scale of Pchatbot is significantly larger than existing Chinese datasets, which might benefit the data-driven models. Besides, current dialogue datasets for personalized chatbot usually contain several persona sentences or attributes. Different from existing datasets, Pchatbot provides anonymized user IDs and timestamps for both posts and responses. This enables the development of personalized dialogue models that directly learn implicit user personality from the user's dialogue history. Our preliminary experimental study benchmarks several state-of-the-art dialogue models to provide a comparison for future work. The dataset can be publicly accessed at Github: https://github.com/qhjqhj00/Pchatbot.

Search result diversification aims to offer diverse documents that cover as many intents as possible. Most existing implicit diversification approaches model diversity through the similarity of document representation, which is indirect and unnatural. To handle the diversity more precisely, we measure the similarity of documents by their similarity of the intent coverage. Specifically, we build a classifier to judge whether two different documents contain the same intent based on the document's content. Then we construct an intent graph to present the complicated relationship of documents and the query. On the intent graph, documents are connected if they are similar, while the query and the document are gradually connected based on the document selection result. Then we employ graph convolutional networks (GCNs) to update the representation of the query and each document by aggregating its neighbors. By this means, we can obtain the context-aware query representation and the intent-aware document representations through the dynamic intent graph during the document selection process. Furthermore, these representations and intent graph features are fused into diversity features. Combined with the traditional relevance features, we obtain the final ranking score that balances the relevance and the diversity. Experimental results show that this implicit diversification model significantly outperforms all existing implicit diversification methods, and it can even beat the state-of-the-art explicit models.

Sequential recommendation aims at predicting users' preferences based on their historical behaviors. However, this recommendation strategy may not perform well in practice due to the sparsity of the real-world data. In this paper, we propose a novel counterfactual data augmentation framework to mitigate the impact of the imperfect training data and empower sequential recommendation models. Our framework is composed of a sampler model and an anchor model. The sampler model aims to generate new user behavior sequences based on the observed ones, while the anchor model is leveraged to provide the final recommendation list, which is trained based on both observed and generated sequences. We design the sampler model to answer the key counterfactual question: "what would a user like to buy if her previously purchased items had been different?". Beyond heuristic intervention methods, we leverage two learning-based methods to implement the sampler model, and thus, improve the quality of the generated sequences when training the anchor model. Additionally, we analyze the influence of the generated sequences on the anchor model in theory and achieve a trade-off between the information and the noise introduced by the generated sequences. Experiments on nine real-world datasets demonstrate our framework's effectiveness and generality.

The user feedbacks could be delayed in many streaming recommendation scenarios. As an example, the user feedbacks to a recommended coupon consist of the immediate feedback on the click event and the delayed feedback on the resultant conversion. The delayed feedbacks pose a challenge of training recommendation models using instances with incomplete labels. When being applied to real products, the challenge becomes more severe as the streaming recommendation models need to be retrained very frequently and the training instances need to be collected over very short time scales. Existing approaches either simply ignore the unobserved feedbacks or heuristically adjust the feedbacks on a static instance set, resulting in biases in the training data and hurting the accuracy of the learned recommenders. In this paper, we propose a novel and theoretic sound counterfactual approach to adjusting the user feedbacks and learning the recommendation models, called CBDF (Counterfactual Bandit with Delayed Feedback). CBDF formulates the streaming recommendation with delayed feedback as a problem of sequential decision making and models it with a batched bandit. To deal with the issue of delayed feedback, at each iteration (episode), a counterfactual importance sampling model is employed to re-weight the original feedbacks and generate the modified rewards. Based on the modified rewards, a batched bandit is learned for conducting online recommendation at the next iteration. Theoretical analysis showed that the modified rewards are statistically unbiased, and the learned bandit policy enjoys a sub-linear regret bound. Experimental results demonstrated that CBDF can outperform the state-of-the-art baselines on a synthetic dataset, the Criteo dataset, and a dataset from Tencent's WeChat app.

The key to personalized search is to build the user profile based on historical behaviour. To deal with the users who lack historical data, group based personalized models were proposed to incorporate the profiles of similar users when re-ranking the results. However, similar users are mostly found based on simple lexical or topical similarity in search behaviours. In this paper, we propose a neural network enhanced method to highlight similar users in semantic space. Furthermore, we argue that the behaviour-based similar users are still insufficient to understand a new query when user's historical activities are limited. To tackle this issue, we introduce the friend network into personalized search to determine the closeness between users in another way. Since the friendship is often formed based on similar background or interest, there are plenty of personalized signals hidden in the friend network naturally. Specifically, we propose a friend network enhanced personalized search model, which groups the user into multiple friend circles based on search behaviours and friend relations respectively. These two types of friend circles are complementary to construct a more comprehensive group profile for refining the personalization. Experimental results show the significant improvement of our model over existing personalized search models.

Existing few-shot learning (FSL) methods make the implicit assumption that the few target class samples are from the same domain as the source class samples. However, in practice, this assumption is often invalid -the target classes could come from a different domain. This poses an additional challenge of domain adaptation (DA) with few training samples. In this paper, the problem of domain-adaptive few-shot learning (DA-FSL) is tackled, which is expected to have wide use in real-world scenarios and requires solving FSL and DA in a unified framework. To this end, we propose a novel domain-adversarial prototypical network (DAPN) model. It is designed to address a specific challenge in DA-FSL: the DA objective means that the source and target data distributions need to be aligned, typically through a shared domain-adaptive feature embedding space; but the FSL objective dictates that the target domain per class distribution must be different from that of any source domain class, meaning aligning the distributions across domains may harm the FSL performance. How to achieve global domain distribution alignment whilst maintaining source/target per-class discriminativeness thus becomes the key. Our solution is to explicitly enhance the source/target per-class separation before domain-adaptive feature embedding learning, to alleviate the negative effect of domain alignment on FSL. Extensive experiments show that our DAPN outperforms the state-of-the-arts. The code is available at https://github.com/dingmyu/DAPN.

Multi-hop Knowledge Base Question Answering (KBQA) aims to find the answer entities that are multiple hops away in the Knowl- edge Base (KB) from the entities in the question. A major challenge is the lack of supervision signals at intermediate steps. Therefore, multi-hop KBQA algorithms can only receive the feedback from the final answer, which makes the learning unstable or ineffective. To address this challenge, we propose a novel teacher-student approach for the multi-hop KBQA task. In our approach, the stu- dent network aims to find the correct answer to the query, while the teacher network tries to learn intermediate supervision signals for improving the reasoning capacity of the student network. The major novelty lies in the design of the teacher network, where we utilize both forward and backward reasoning to enhance the learning of intermediate entity distributions. By considering bidi- rectional reasoning, the teacher network can produce more reliable intermediate supervision signals, which can alleviate the issue of spurious reasoning. Extensive experiments on three benchmark datasets have demonstrated the effectiveness of our approach on the KBQA task.

This paper concerns user preference estimation in multi-round conversational recommender systems (CRS), which interacts with users by asking questions about attributes and recommending items multiple times in one conversation. Multi-round CRS such as EAR have been proposed in which the user's online feedback at both attribute level and item level can be utilized to estimate user preference and make recommendations. Though preliminary success has been shown, existing user preference models in CRS usually use the online feedback information as independent features or training instances, overlooking the relation between attribute-level and item-level feedback signals. The relation can be used to more precisely identify the reasons (e.g., some certain attributes) that trigger the rejection of an item, leading to more fine-grained utilization of the feedback information. To address aforementioned issue, this paper proposes a novel preference estimation model tailored for multi-round CRS, called Feedback-guided Preference Adaptation Network (FPAN). In FPAN, two gating modules are designed to respectively adapt the original user embedding and item-level feedback, both according to the online attribute-level feedback. The gating modules utilize the fine-grained attribute-level feedback to revise the user embedding and coarse-grained item-level feedback, achieving more accurate user preference estimation by considering the relation between feedback. Experimental results on two benchmarks showed that FPAN outperformed the state-of-the-art user preference models in CRS, and the multi-round CRS can also be enhanced by using FPAN as its recommender component.

The evolution of mobile apps has greatly changed the way that we live. It becomes increasingly important to understand and model the users on mobile apps. Instead of focusing on some specific app alone, it has become a popular paradigm to study the user behavior on various mobile apps in a symbiotic environment.

Personalized search returns each user more accurate results by collecting the user’s historical search behaviors to infer her interests and query intents. However, it brings the risk of user privacy leakage, and this may greatly limit the practical application of personalized search. In this paper, we focus on the problem of privacy protection in personalized search, and propose a privacy protection enhanced personalized search framework, denoted with FedPS. Under this framework, we keep each user’s private data on her individual client, and train a shared personalized ranking model with all users’ decentralized data by means of federated learning. We implement two models within the framework: the first one applies a personalization model with a personal module that fits the user’s data distribution to alleviate the challenge of data heterogeneity in federated learning; the second model introduces trustworthy proxies and group servers to solve the problems of limited communication, performance bottleneck and privacy attack for FedPS. Experimental results verify that our proposed framework can enhance privacy protection without losing too much accuracy.

Zero-shot learning (ZSL) is made possible by learning a projection function between a feature space and a semantic space (e.g., an attribute space). Key to ZSL is thus to learn a projection that is robust against the often large domain gap between the seen and unseen class domains. In this work, this is achieved by unseen class data synthesis and robust projection function learning. Specifically, a novel semantic data synthesis strategy is proposed, by which semantic class prototypes (e.g., attribute vectors) are used to simply perturb seen class data for generating unseen class ones. As in any data synthesis/hallucination approach, there are ambiguities and uncertainties on how well the synthesised data can capture the targeted unseen class data distribution. To cope with this, the second contribution of this work is a novel projection learning model termed competitive bidirectional projection learning (BPL) designed to best utilise the ambiguous synthesised data. Specifically, we assume that each synthesised data point can belong to any unseen class; and the most likely two class candidates are exploited to learn a robust projection function in a competitive fashion. As a third contribution, we show that the proposed ZSL model can be easily extended to few-shot learning (FSL) by again exploiting semantic (class prototype guided) feature synthesis and competitive BPL. Extensive experiments show that our model achieves the state-of-the-art results on both problems.

Conversational systems now attract great attention due to their promising potential and commercial values. To build a conversational system with moderate intelligence is challenging and requires big (conversational) data, as well as interdisciplinary techniques. Thanks to the prosperity of the Web, the massive data available greatly facilitate data-driven methods such as deep learning for human-computer conversational systems. In general, retrieval-based conversational systems apply various matching schema between query utterances and responses, but the classic retrieval paradigm suffers from prominent weakness for conversations: the system finds similar responses given a particular query. For real human-to-human conversations, on the contrary, responses can be greatly different yet all are possibly appropriate. The observation reveals the diversity phenomenon in conversations. In this article, we ascribe the lack of conversational diversity to the reason that the query utterances are statically modeled regardless of candidate responses through traditional methods. To this end, we propose a dynamic representation learning strategy that models the query utterances and different response candidates in an interactive way. To be more specific, we propose a Respond-with-Diversity model augmented by the memory module interacting with both the query utterances and multiple candidate responses. Hence, we obtain dynamic representations for the input queries conditioned on different response candidates. We frame the model as an end-to-end learnable neural network. In the experiments, we demonstrate the effectiveness of the proposed model by achieving a good appropriateness score and much better diversity in retrieval-based conversations between humans and computers.

Personalized search is a promising way to improve search qualities by taking user interests into consideration. Recently, machine learning and deep learning techniques have been successfully applied to search result personalization. Most existing models simply regard the personal search history as a static set of user behaviors and learn fixed ranking strategies based on all the recorded data. Though improvements have been achieved, the essence that the search process is a sequence of interactions between the search engine and user is ignored. The user’s interests may dynamically change during the search process, therefore, it would be more helpful if a personalized search model could track the whole interaction process and adjust its ranking strategy continuously. In this article, we adapt reinforcement learning to personalized search and propose a framework, referred to as RLPS. It utilizes a Markov Decision Process ( MDP ) to track sequential interactions between the user and search engine, and continuously update the underlying personalized ranking model with the user’s real-time feedback to learn the user’s dynamic interests. Within this framework, we implement two models: the listwise RLPS-L and the hierarchical RLPS-H. RLPS-L interacts with users and trains the ranking model with document lists, while RLPS-H improves model training by designing a layered structure and introducing document pairs. In addition, we also design a feedback-aware personalized ranking component to capture the user’s feedback, which impacts the user interest profile for the next query. Significant improvements over existing personalized search models are observed in the experiments on the public AOL search log and a commercial log.

Most state-of-the-art named entity recognition systems are designed to process each sentence within a document independently. These systems are easy to confuse entity types when the context information in a sentence is not sufficient enough. To utilize the context information within the whole document, most document-level work let neural networks on their own to learn the relation across sentences, which is not intuitive enough for us humans. In this paper, we divide entities to multi-token entities that contain multiple tokens and single-token entities that are composed of a single token. We propose that the context information of multi-token entities should be more reliable in document-level NER for news articles. We design a fusion attention mechanism which not only learns the semantic relevance between occurrences of the same token, but also focuses more on occurrences belonging to multi-tokens entities. To identify multi-token entities, we design an auxiliary task namely ‘Multi-token Entity Classification’ and perform this task simultaneously with document-level NER. This auxiliary task is simplified from NER and doesn't require extra annotation. Experimental results on the CoNLL-2003 dataset and OntoNotesnbm dataset show that our model outperforms state-of-the-art sentence-level and document-level NER methods.

With the ever-increasing growth of online recruitment data, job-resume matching has become an important task to automatically match jobs with suitable resumes. This task is typically casted as a supervised text matching problem. Supervised learning is powerful when the labeled data is sufficient. However, on online recruitment platforms, job-resume interaction data is sparse and noisy, which affects the performance of job-resume match algorithms.

Personalized review generation (PRG) aims to automatically produce review text reflecting user preference, which is a challenging natural language generation task. Most of previous studies do not explicitly model factual description of products, tending to generate uninformative content. Moreover, they mainly focus on word-level generation, but cannot accurately reflect more abstractive user preference in multiple aspects. To address the above issues, we propose a novel knowledgeenhanced PRG model based on capsule graph neural network (CapsGNN). We first construct a heterogeneous knowledge graph (HKG) for utilizing rich item attributes. We adopt Caps-GNN to learn graph capsules for encoding underlying characteristics from the HKG. Our generation process contains two major steps, namely aspect sequence generation and sentence generation. First, based on graph capsules, we adaptively learn aspect capsules for inferring the aspect sequence. Then, conditioned on the inferred aspect label, we design a graph-based copy mechanism to generate sentences by incorporating related entities or words from HKG. To our knowledge, we are the first to utilize knowledge graph for the PRG task. The incorporated KG information is able to enhance user preference at both aspect and word levels. Extensive experiments on three real-world datasets have demonstrated the effectiveness of our model on the PRG task.

Personalized search aims to improve the search quality by re-ranking the candidate document list based on user's historical behavior. Existing approaches focus on modeling the order information of user's search history by sequential methods such as Recurrent Neural Network (RNN). However, these methods usually ignore the fine-grained time information associated with user actions. In fact, the time intervals between queries can help to capture the evolution of query intent and document interest of users. Besides, the time intervals between past actions and current query can reflect the re-finding tendency more accurately than discrete steps in RNN. In this paper, we propose PSTIE, a fine-grained Time Information Enhanced model to construct more accurate user interest representations for Personalized Search. To capture the short-term interest of users, we design time-aware LSTM architectures for modeling the subtle interest evolution of users in continuous time. We further leverage time in calculating the re-finding possibility of users to capture the long-term user interest. We propose two methods to utilize the time-enhanced user interest into personalized ranking. Experiments on two datasets show that PSTIE can effectively improve the ranking quality over state-of-the-art models.

Search results returned by search engines need to be diversified in order to satisfy different information needs of different users. Several supervised learning models have been proposed for diversifying search results in recent years. Most of the existing supervised methods greedily compare each candidate document with the selected document sequence and select the next local optimal document. However, the information utility of each candidate document is not independent with each other, and research has shown that the selection of a candidate document will affect the utilities of other candidate documents. As a result, the local optimal document rankings will not lead to the global optimal rankings. In this paper, we propose a new supervised diversification framework to address this issue. Based on a self-attention encoder-decoder structure, the model can take the whole candidate document sequence as input, and simultaneously leverage both the novelty and the subtopic coverage of the candidate documents. We call this framework Diversity Encoder with Self-Attention (DESA). Comparing with existing supervised methods, this framework can model the interactions between all candidate documents and return their diversification scores based on the whole candidate document sequence. Experimental results show that our proposed framework outperforms existing methods. These results confirm the effectiveness of modeling all the candidate documents for the overall novelty and subtopic coverage globally, instead of comparing every single candidate document with the selected sequence document selection.

Top-N item recommendation has been a widely studied task from implicit feedback. Although much progress has been made with neural methods, there is increasing concern on appropriate evaluation of recommendation algorithms. In this paper, we revisit alternative experimental settings for evaluating top-N recommendation algorithms, considering three important factors, namely dataset splitting, sampled metrics and domain selection. We select eight representative recommendation algorithms (covering both traditional and neural methods) and construct extensive experiments on a very large dataset. By carefully revisiting different options, we make several important findings on the three factors, which directly provide useful suggestions on how to appropriately set up the experiments for top-N item recommendation.

Recently, significant progress has been made in sequential recommendation with deep learning. Existing neural sequential recommendation models usually rely on the item prediction loss to learn model parameters or data representations. However, the model trained with this loss is prone to suffer from data sparsity problem. Since it overemphasizes the final performance, the association or fusion between context data and sequence data has not been well captured and utilized for sequential recommendation.

Recently, conversational recommender system (CRS) has become an emerging and practical research topic. Most of the existing CRS methods focus on learning effective preference representations for users from conversation data alone. While, we take a new perspective to leverage historical interaction data for improving CRS. For this purpose, we propose a novel pre-training approach to integrating both item-based preference sequence (from historical interaction data) and attribute-based preference sequence (from conversation data) via pre-training methods. We carefully design two pre-training tasks to enhance information fusion between item- and attribute-based preference. To improve the learning performance, we further develop an effective negative sample generator which can produce high-quality negative samples. Experiment results on two real-world datasets have demonstrated the effectiveness of our approach for improving CRS.

Conversational recommender systems (CRS) aim to recommend high-quality items to users through interactive conversations.To develop an effective CRS, the support of high-quality datasets is essential.Existing CRS datasets mainly focus on immediate requests from users, while lack proactive guidance to the recommendation scenario.In this paper, we contribute a new CRS dataset named TG-ReDial (Recommendation through Topic-Guided Dialog).Our dataset has two major features.First, it incorporates topic threads to enforce natural semantic transitions towards the recommendation scenario.Second, it is created in a semi-automatic way, hence human annotation is more reasonable and controllable.Based on TG-ReDial, we present the task of topic-guided conversational recommendation, and propose an effective approach to this task.Extensive experiments have demonstrated the effectiveness of our approach on three sub-tasks, namely topic prediction, item recommendation and response generation.TG-ReDial is available at https

One approach to matching texts from asymmetrical domains is projecting the input sequences into a common semantic space as feature vectors upon which the matching function can be readily defined and learned.In realworld matching practices, it is often observed that with the training goes on, the feature vectors projected from different domains tend to be indistinguishable.The phenomenon, however, is often overlooked in existing matching models.As a result, the feature vectors are constructed without any regularization, which inevitably increases the difficulty of learning the downstream matching functions.In this paper, we propose a novel match method tailored for text matching in asymmetrical domains, called WD-Match.In WD-Match, a Wasserstein distance-based regularizer is defined to regularize the features vectors projected from different domains.As a result, the method enforces the feature projection function to generate vectors such that those correspond to different domains cannot be easily discriminated.The training process of WD-Match amounts to a game that minimizes the matching loss regularized by the Wasserstein distance.WD-Match can be used to improve different text matching methods, by using the method as its underlying matching model.Four popular text matching methods have been exploited in the paper.Experimental results based on four publicly available benchmarks showed that WD-Match consistently outperformed the underlying methods and the baselines.

Conversational recommender systems (CRS) aim to recommend high-quality items to users through interactive conversations. Although several efforts have been made for CRS, two major issues still remain to be solved. First, the conversation data itself lacks of sufficient contextual information for accurately understanding users' preference. Second, there is a semantic gap between natural language expression and item-level user preference.

Graph Neural Networks (GNN) are an emerging field for learning on non-Euclidean data. Recently, there has been increased interest in designing GNN that scales to large graphs. Most existing methods use "graph sampling" or "layer-wise sampling" techniques to reduce training time; However, these methods still suffer from degrading performance and scalability problems when applying to graphs with billions of edges. In this paper, we present GBP, a scalable GNN that utilizes a localized bidirectional propagation process from both the feature vector and the training/testing nodes. Theoretical analysis shows that GBP is the first method that achieves sub-linear time complexity for both the precomputation and the training phases. An extensive empirical study demonstrates that GBP achieves state-of-the-art performance with significantly less training/testing time. Most notably, GBP is able to deliver superior performance on a graph with over 60 million nodes and 1.8 billion edges in less than 2,000 seconds on a single machine.

Most commercial industrial recommender systems have built their closed feedback loops. Though it is helpful in item recommendation and model training, the closed feedback loop may lead to the so-called bias problems, including the position bias, selection bias and popularity bias. The recommendation models trained with biased may hurt the user experiences by recommending homogenous items. How to control the biases in the closed feedback loop has become one of major challenges in modern recommender systems. This talk discusses the counterfactual learning technologies for tackling the bias problem in recommendation.

Search result diversification aims to retrieve diverse results to cover as many subtopics related to the query as possible. Recent studies showed that supervised diversification models are able to outperform the heuristic approaches, by automatically learning a diversification function other than using manually designed score functions. The main challenge of training a diversification model is the lack of high-quality training samples. Due to the involvement of dependence between documents in the ranker, it is very hard for training algorithms to select effective positive and negative ranking lists to train a reliable ranking model, given a large number of candidate documents within which different documents are relevant to different subtopics. To tackle this problem, we propose a supervised diversification framework based on Generative Adversarial Network (GAN). It consists of a generator and a discriminator interacting with each other in a minimax game. Specifically, the generator generates more confusing negative samples for the discriminator, and the discriminator sends back complementary ranking signals to the generator. Furthermore, we explicitly exploit subtopics in the generator, whereas focusing on modeling document similarity in the discriminator. Through such a minimax game, we are able to obtain better ranking models by combining ranking signals learned by the generator and the discriminator. Experimental results on the TREC Web Track dataset show that the proposed method can significantly outperform existing diversification methods.

This paper presents a knowledge graph enhanced personalized search model, KEPS. For each user and her queries, KEPS first con- ducts personalized entity linking on the queries and forms better intent representations; then it builds a knowledge enhanced profile for the user, using memory networks to store the predicted search intents and linked entities in her search history. The knowledge enhanced user profile and intent representation are then utilized by KEPS for better, knowledge enhanced, personalized search. Furthermore, after providing personalized search for each query, KEPS leverages user's feedback (click on documents) to post-adjust the entity linking on previous queries. This fixes previous linking errors and improves ranking quality for future queries. Experiments on the public AOL search log demonstrate the advantage of knowledge in personalized search: personalized entity linking better reflects user's search intent, the memory networks better maintain user's subtle preferences, and the post linking adjustment fixes some linking errors with the received feedback signals. The three components together lead to a significantly better ranking accuracy of KEPS.

In learning-to-rank for information retrieval, a ranking model is automatically learned from the data and then utilized to rank the sets of retrieved documents. Therefore, an ideal ranking model would be a mapping from a document set to a permutation on the set, and should satisfy two critical requirements: (1) it should have the ability to model cross-document interactions so as to capture local context information in a query; (2) it should be permutation-invariant, which means that any permutation of the inputted documents would not change the output ranking. Previous studies on learning-to-rank either design uni-variate scoring functions that score each document separately, and thus failed to model the cross-document interactions; or construct multivariate scoring functions that score documents sequentially, which inevitably sacrifice the permutation invariance requirement. In this paper, we propose a neural learning-to-rank model called SetRank which directly learns a permutation-invariant ranking model defined on document sets of any size. SetRank employs a stack of (induced) multi-head self attention blocks as its key component for learning the embeddings for all of the retrieved documents jointly. The self-attention mechanism not only helps SetRank to capture the local context information from cross-document interactions, but also to learn permutation-equivariant representations for the inputted documents, which therefore achieving a permutation-invariant ranking model. Experimental results on three benchmarks showed that the SetRank significantly outperformed the baselines include the traditional learning-to-rank models and state-of-the-art Neural IR models.

Recently, deep learning has made significant progress in the task of sequential recommendation. Existing neural sequential recommenders typically adopt a generative way trained with Maximum Likelihood Estimation (MLE). When context information (called factor) is involved, it is difficult to analyze when and how each individual factor would affect the final recommendation performance.

This paper concerns reinforcement learning~(RL) of the document ranking models for information retrieval~(IR). One branch of the RL approaches to ranking formalize the process of ranking with Markov decision process~(MDP) and determine the model parameters with policy gradient. Though preliminary success has been shown, these approaches are still far from achieving their full potentials. Existing policy gradient methods directly utilize the absolute performance scores (returns) of the sampled document lists in its gradient estimations, which may cause two limitations: 1) fail to reflect the relative goodness of documents within the same query, which usually is close to the nature of IR ranking; 2) generate high variance gradient estimations, resulting in slow learning speed and low ranking accuracy. To deal with the issues, we propose a novel policy gradient algorithm in which the gradients are determined using pairwise comparisons of two document lists sampled within the same query. The algorithm, referred to as Pairwise Policy Gradient (PPG), repeatedly samples pairs of document lists, estimates the gradients with pairwise comparisons, and finally updates the model parameters. Theoretical analysis shows that PPG makes an unbiased and low variance gradient estimations. Experimental results have demonstrated performance gains over the state-of-the-art baselines in search result diversification and text retrieval.

Personalized search is a task to tailor the general document ranking list based on user interests to better satisfy the user's information need. Many personalized search models have been proposed and demonstrated their capability to improve search quality. The general idea of most approaches is to build a user interest profile according to the user's search history, and then re-rank the documents based on the matching scores between the created user profile and candidate documents. In this paper, we propose to solve the problem of personalized search in an alternative way. We know that there are many ambiguous words in natural language such as 'Apple', and people with different knowledge backgrounds and interests have personalized understandings of these words. Therefore, for different users, such a word should own different semantic representations. Motivated by this idea, we design a personalized search model based on personal word embeddings, referred to as PEPS. Specifically, we train personal word embeddings for each user in which the representation of each word is mainly decided by the user's personal data. Then, we obtain the personalized word and contextual representations of the query and documents with an attention function. Finally, we use a matching model to calculate the matching score between the personalized query and document representations. Experiments on two datasets verify that our model can significantly improve state-of-the-art personalization models.

The key to personalized search is to clarify the meaning of current query based on user's search history. Previous personalized studies tried to build user profiles on the basis of historical data to tailor the ranking. However, we argue that the user profile based methods do not really disambiguate the current query. They still retain some semantic bias when building user profiles. In this paper, we propose to encode history with context-aware representation learning to enhance the representation of current query, which is a direct way to clarify the user's information need. Specifically, endowed with the benefit from transformer on aggregating contextual information, we devise a query disambiguation model to parse the meaning of current query in multiple stages. Moreover, for covering the cases that current query is not sufficient to express the intent, we train a personalized language model to predict user intent from existing queries. Under the interaction of two sub-models, we can generate the context-aware representation of current query and re-rank the results based on it. Experimental results show the significant improvement of our model compared with previous methods.

SimRank is a popular measurement for evaluating the node-to-node similarities based on the graph topology. In recent years, single-source and top-k SimRank queries have received increasing attention due to their applications in web mining, social network analysis, and spam detection. However, a fundamental obstacle in studying SimRank has been the lack of ground truths. The only exact algorithm, Power Method, is computationally infeasible on graphs with more than 106 nodes. Consequently, no existing work has evaluated the actual trade-offs between query time and accuracy on large real-world graphs. In this paper, we present ExSim, the first algorithm that computes the exact single-source and top-k SimRank results on large graphs. With high probability, this algorithm produces ground truths with a rigorous theoretical guarantee. We conduct extensive experiments on real-world datasets to demonstrate the efficiency of ExactSim. The results show that ExactSim provides the ground truth for any single-source SimRank query with a precision up to 7 decimal places within a reasonable query time.

The goal of personalized search is to tailor the document ranking list to meet user's individual needs. Previous studies showed users usually look for the information that has been searched before. This is called re-finding behavior which is widely explored in existing personalized search approaches. However, most existing methods for identifying re-finding behavior focus on simple lexical similarities between queries. In this paper, we propose to construct memory networks (MN) to support the identification of more complex re-finding behavior. Specifically, incorporating semantic information, we devise two external memories to make an expansion of re-finding based on the query and the document respectively. We further design an intent memory to recognize session-based re-finding behavior. Endowed with these memory networks, we can build a fine-grained user model dynamically based on the current query and documents, and use the model to re-rank the results. Experimental results show the significant improvement of our model compared with traditional methods.

The task of Knowledge Graph Completion (KGC) aims to automatically infer the missing fact information in Knowledge Graph (KG). In this paper, we take a new perspective that aims to leverage rich user-item interaction data (user interaction data for short) for improving the KGC task. Our work is inspired by the observation that many KG entities correspond to online items in application systems. However, the two kinds of data sources have very different intrinsic characteristics, and it is likely to hurt the original performance using simple fusion strategy.

Personalized search improves generic ranking models by taking user interests into consideration and returning more accurate search results to individual users. In recent years, machine learning and deep learning techniques have been successfully applied in personalized search. Most existing personalization models simply regard the search history as a static set of user behaviours and learn fixed ranking strategies based on the recorded data. Though improvements have been observed, it is obvious that these methods ignore the dynamic nature of the search process: search is a sequence of interactions between the search engine and the user. During the search process, the user interests may dynamically change. It would be more helpful if a personalized search model could track the whole interaction process and update its ranking strategy continuously. In this paper, we propose a reinforcement learning based personalization model, referred to as RLPer, to track the sequential interactions between the users and search engine with a hierarchical Markov Decision Process (MDP). In RLPer, the search engine interacts with the user to update the underlying ranking model continuously with real-time feedback. And we design a feedback-aware personalized ranking component to catch the user’s feedback which has impacts on the user interest profile for the next query. Experimental results on the publicly available AOL search log verify that our proposed model can significantly outperform state-of-the-art personalized search models.

This article presents a network embedding approach to automatically generate tags for microblog users. Instead of using text data, we aim to annotate microblog users with meaningful tags by leveraging rich social link data. To utilize directed social links, we use two kinds of node representations for modeling user interest in terms of their followers and followees, respectively. To alleviate the sparsity problem, we propose a novel method based on two transformation functions for capturing implicit interest similarity. Different from previous works on capturing high-order proximity, our model is able to directly characterize the effect of the context user on the proximity of node pairs. Another novelty of our model is that the importance scores of users learned from the classic PageRank algorithm are utilized to set the link weights. By using such weights, our model is more capable of disentangling the interest similarity evidence of a link. We jointly consider the above factors when designing the final objective function. We construct a very large evaluation set consisting of 2.6M users, 0.5M tags, and 0.8B following links. To our knowledge, it is the largest reported dataset for microblog user tagging in the literature. Extensive experiments on this dataset demonstrate the effectiveness of the proposed approach. We implement this approach with several optimization techniques, which makes our model easy to scale to very large social networks. Ubiquitous social links provide important data resources to understand user interests. Our work provides an effective and efficient solution to annotate user interests solely using the link data, which has important practical value in industry. To illustrate the use of our models, we implement a demonstration system for visualizing, navigating, and searching microblog users.

Generating long and informative review text is a challenging natural language generation task.Previous work focuses on word-level generation, neglecting the importance of topical and syntactic characteristics from natural languages.In this paper, we propose a novel review generation model by characterizing an elaborately designed aspect-aware coarse-tofine generation process.First, we model the aspect transitions to capture the overall content flow.Then, to generate a sentence, an aspectaware sketch will be predicted using an aspectaware decoder.Finally, another decoder fills in the semantic slots by generating corresponding words.Our approach is able to jointly utilize aspect semantics, syntactic sketch, and context information.Extensive experiments results have demonstrated the effectiveness of the proposed model.

Automated deception detection (ADD) from real-life videos is a challenging task. It specifically needs to address two problems: (1) Both face and body contain useful cues regarding whether a subject is deceptive. How to effectively fuse the two is thus key to the effectiveness of an ADD model. (2) Real-life deceptive samples are hard to collect; learning with limited training data thus challenges most deep learning based ADD models. In this work, both problems are addressed. Specifically, for face-body multimodal learning, a novel face-focused cross-stream network (FFCSN) is proposed. It differs significantly from the popular two-stream networks in that: (a) face detection is added into the spatial stream to capture the facial expressions explicitly, and (b) correlation learning is performed across the spatial and temporal streams for joint deep feature learning across both face and body. To address the training data scarcity problem, our FFCSN model is trained with both meta learning and adversarial learning. Extensive experiments show that our FFCSN model achieves state-of-the-art results. Further, the proposed FFCSN model as well as its robust training strategy are shown to be generally applicable to other human-centric video analysis tasks such as emotion recognition from user-generated videos.

Visual dialog is a challenging vision-language task, which requires the agent to answer multi-round questions about an image. It typically needs to address two major problems: (1) How to answer visually-grounded questions, which is the core challenge in visual question answering (VQA); (2) How to infer the co-reference between questions and the dialog history. An example of visual co-reference is: pronouns (e.g., “they”) in the question (e.g., “Are they on or off?”) are linked with nouns (e.g., “lamps”) appearing in the dialog history (e.g., “How many lamps are there?”) and the object grounded in the image. In this work, to resolve the visual co-reference for visual dialog, we propose a novel attention mechanism called Recursive Visual Attention (RvA). Specifically, our dialog agent browses the dialog history until the agent has sufficient confidence in the visual co-reference resolution, and refines the visual attention recursively. The quantitative and qualitative experimental results on the large-scale VisDial v0.9 and v1.0 datasets demonstrate that the proposed RvA not only outperforms the state-of-the-art methods, but also achieves reasonable recursion and interpretable attention maps without additional annotations. The code is available at https://github.com/yuleiniu/rva.

Shuqing Bian, Wayne Xin Zhao, Yang Song, Tao Zhang, Ji-Rong Wen. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Siqing Li, Wayne Xin Zhao, Eddy Jing Yin, Ji-Rong Wen. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

We bring the novel idea of exploiting motifs into network embedding, in a dual-level network representation learning model called RUM (network Representation learning Using Motifs). Towards the leveraging of graph motifs that constitute higher-order organizations in a network, we propose two strategies, namely MotifWalk and MotifRe-weighting for learning motif-aware network embeddings. Motif-based and node-based representations are simultaneously generated, so that both the high-order structures and each node's individual properties are preserved in the final embeddings. We demonstrate that RUM has strong and well-balanced capability of preserving lowerorder proximities while discovering and capturing higher-order network structures. In empirical evaluation, RUM is tested on multiple public datasets, that range from small to medium citation networks to a large social network with more than a million nodes. Results show that the use of motifs in the representation learning process brings substantial benefits in reallife tasks, resulting in up to 12% microF1 and 8% macroF1 relative gains for node classification performance over the bestperforming competing methods.

Zero-shot learning (ZSL) is originally designed to address the small sample size problem often encountered in computer vision by recognizing unseen object classes without any training samples. Existing ZSL models (particularly deep ones) often assume that hundreds of labelled samples are collected from each seen class. In real-world applications, this assumption tends to become invalid. Therefore, a new ZSL setting is concerned in this paper: each seen class only has few labelled samples, while each unseen class still has no samples. This is more challenging yet more useful/practical than the conventional ZSL setting. To overcome the extreme label scarcity, we choose to obtain more training samples from image search engine for data augmentation: the name of each seen class is used as the query of Google, and the top returned images can be viewed as the noisy labelled samples for this seen class. With the augmented but noisy labelled training data, a novel inductive ZSL model is proposed by formulating label noise reduction (LNR) and semantic projection learning (SPL) within a unified framework: (1) LNR aims to refine the noisy labelled samples for projection learning; (2) SPL aims to learn the projection function with the refined training data. Extensive experiments show that our ZSL model outperforms the state-of-the-art alternatives.

In e-commerce, users' demands are not only conditioned by their profile and preferences, but also by their recent purchases that may generate new demands, as well as periodical demands that depend on purchases made some time ago. We call them respectively short-term demands and long-term demands. In this paper, we propose a novel self-attentive Continuous-Time Recommendation model (CTRec) for capturing the evolving demands of users over time. For modeling such time-sensitive demands, a Demand-aware Hawkes Process (DHP) framework is designed in CTRec to learn from the discrete purchase records of users. More specifically, a convolutional neural network is utilized to capture the short-term demands; and a self-attention mechanism is employed to capture the periodical purchase cycles of long-term demands. All types of demands are fused in DHP to make final continuous-time recommendations. We conduct extensive experiments on four real-world commercial datasets to demonstrate that CTRec is effective for general sequential recommendation problems, including next-item and next-session/basket recommendations. We observe in particular that CTRec is capable of learning the purchase cycles of products and estimating the purchase time of a product given a user.

Fine-grained image classification and retrieval become topical in both computer vision and information retrieval. In real-life scenarios, fine-grained tasks tend to appear along with coarse-grained tasks when the observed object is coming closer. However, in previous works, the combination of fine-grained and coarse-grained tasks was often ignored. In this paper, we define a new problem called coarse-to-fine grained classification (C2FGC) which aims to recognize the classes of objects in multiple resolutions (from low to high). To solve this problem, we propose a novel Multi-linear Pooling with Hierarchy (MLPH) model. Specifically, we first design a multi-linear pooling module to include both trilinear and bilinear pooling, and then formulate the coarse-grained and fine-grained tasks within a unified framework. Experiments on two benchmark datasets show that our model achieves state-of-the-art results.

In this paper we propose a novel reinforcement learning based model for named entity recognition (NER), referred to as MM-NER. Inspired by the methodology of the AlphaGo Zero, MM-NER formalizes the problem of named entity recognition with a Monte-Carlo tree search (MCTS) enhanced Markov decision process (MDP) model, in which the time steps correspond to the positions of words in a sentence from left to right, and each action corresponds to assign an NER tag to a word. Two Gated Recurrent Units (GRU) are used to summarize the past tag assignments and words in the sentence. Based on the outputs of GRUs, the policy for guiding the tag assignment and the value for predicting the whole tagging accuracy of the whole sentence are produced. The policy and value are then strengthened with MCTS, which takes the produced raw policy and value as inputs, simulates and evaluates the possible tag assignments at the subsequent positions, and outputs a better search policy for assigning tags. A reinforcement learning algorithm is proposed to train the model parameters. Empirically, we show that MM-NER can accurately predict the tags thanks to the exploratory decision making mechanism introduced by MCTS. It outperformed the conventional sequence tagging baselines and performed equally well with the state-of-the-art baseline BLSTM-CRF.

Personalized search aims to adapt document ranking to user's personal interests. Traditionally, this is done by extracting click and topical features from historical data in order to construct a user profile. In recent years, deep learning has been successfully used in personalized search due to its ability of automatic feature learning. However, the small amount of noisy personal data poses challenges to deep learning models to learn the personalized classification boundary between relevant and irrelevant results. In this paper, we propose PSGAN, a Generative Adversarial Network (GAN) framework for personalized search. By means of adversarial training, we enforce the model to pay more attention to training data that are difficult to distinguish. We use the discriminator to evaluate personalized relevance of documents and use the generator to learn the distribution of relevant documents. Two alternative ways to construct the generator in the framework are tested: based on the current query or based on a set of generated queries. Experiments on data from a commercial search engine show that our models can yield significant improvements over state-of-the-art models.

SimRank is a classic measure of the similarities of nodes in a graph. Given a node u in graph $G =(V, E)$, a \em single-source SimRank query returns the SimRank similarities $s(u, v)$ between node u and each node $v \in V$. This type of queries has numerous applications in web search and social networks analysis, such as link prediction, web mining, and spam detection. Existing methods for single-source SimRank queries, however, incur query cost at least linear to the number of nodes n, which renders them inapplicable for real-time and interactive analysis. This paper proposes \prsim, an algorithm that exploits the structure of graphs to efficiently answer single-source SimRank queries. \prsim uses an index of size $O(m)$, where m is the number of edges in the graph, and guarantees a query time that depends on the \em reverse PageRank distribution of the input graph. In particular, we prove that \prsim runs in sub-linear time if the degree distribution of the input graph follows the power-law distribution, a property possessed by many real-world graphs. Based on the theoretical analysis, we show that the empirical query time of all existing SimRank algorithms also depends on the reverse PageRank distribution of the graph. Finally, we present the first experimental study that evaluates the absolute errors of various SimRank algorithms on large graphs, and we show that \prsim outperforms the state of the art in terms of query time, accuracy, index size, and scalability.

In this paper, we focus on the task of sequential recommendation using taxonomy data. Existing sequential recommendation methods usually adopt a single vectorized representation for learning the overall sequential characteristics, and have a limited modeling capacity in capturing multi-grained sequential characteristics over context information. Besides, existing methods often directly take the feature vectors derived from context information as auxiliary input, which is difficult to fully exploit the structural patterns in context information for learning preference representations. To address above issues, we propose a novel Taxonomy-aware Multi-hop Reasoning Network, named TMRN, which integrates a basic GRU-based sequential recommender with an elaborately designed memory-based multi-hop reasoning architecture. For enhancing the reasoning capacity, we incorporate taxonomy data as structural knowledge to instruct the learning of our model. We associate the learning of user preference in sequential recommendation with the category hierarchy in the taxonomy. Given a user, for each recommendation, we learn a unique preference representation corresponding to each level in the taxonomy based on her/his overall sequential preference. In this way, the overall, coarse-grained preference representation can be gradually refined in different levels from general to specific, and we are able to capture the evolvement and refinement of user preference over the taxonomy, which makes our model highly explainable. Extensive experiments show that our proposed model is superior to state-of-the-art baselines in terms of both effectiveness and interpretability.

Image annotation aims to annotate a given image with a variable number of class labels corresponding to diverse visual concepts. In this paper, we address two main issues in large-scale image annotation: 1) how to learn a rich feature representation suitable for predicting a diverse set of visual concepts ranging from object, scene to abstract concept and 2) how to annotate an image with the optimal number of class labels. To address the first issue, we propose a novel multi-scale deep model for extracting rich and discriminative features capable of representing a wide range of visual concepts. Specifically, a novel two-branch deep neural network architecture is proposed, which comprises a very deep main network branch and a companion feature fusion network branch designed for fusing the multi-scale features computed from the main branch. The deep model is also made multi-modal by taking noisy user-provided tags as model input to complement the image input. For tackling the second issue, we introduce a label quantity prediction auxiliary task to the main label prediction task to explicitly estimate the optimal label number for a given image. Extensive experiments are carried out on two large-scale image annotation benchmark datasets, and the results show that our method significantly outperforms the state of the art.

Presents corrections to author information from the paper, “Characterizing and predicting early reviewers for effective product marketing on e-commerce websites,” (Bai, T., et al), IEEE Trans. Knowl. Data Eng., vol. 30, no. 12, pp. 2271–2284, Dec. 2018.

Search results personalization has become an effective way to improve the quality of search engines. Previous studies extracted information such as past clicks, user topical interests, query click entropy and so on to tailor the original ranking. However, few studies have taken into account the sequential information underlying previous queries and sessions. Intuitively, the order of issued queries is important in inferring the real user interests. And more recent sessions should provide more reliable personal signals than older sessions. In addition, the previous search history and user behaviors should influence the personalization of the current query depending on their relatedness. To implement these intuitions, in this paper we employ a hierarchical recurrent neural network to exploit such sequential information and automatically generate user profile from historical data. We propose a query-aware attention model to generate a dynamic user profile based on the input query. Significant improvement is observed in the experiment with data from a commercial search engine when compared with several traditional personalization models. Our analysis reveals that the attention model is able to attribute higher weights to more related past sessions after fine training.

We propose and investigate a novel query type named trajectory search by regions of interest (TSR query). Given an argument set of trajectories, a TSR query takes a set of regions of interest as a parameter and returns the trajectory in the argument set with the highest spatial-density correlation to the query regions. This type of query is useful in applications such as trip planning and recommendation. To process the TSR query, a set of new metrics are defined to model spatial-density correlations. An efficient trajectory search algorithm is developed that exploits upper and lower bounds to prune the search space and that adopts a query-source selection strategy, as well as integrates a heuristic search strategy based on priority ranking to schedule multiple query sources. The performance of TSR query processing is studied in extensive experiments based on real and synthetic spatial data.

The denoising and deblurring of images are the two essential restoration tasks in the document image processing task. As the preprocessing stages of the processing pipeline, the quality of denoising and deblurring heavily influences the result of subsequent tasks, such as character detection and recognition. In this paper, we propose a novel neural method for restoring document images. We named our network Skip-Connected Deep Convolutional Autoencoder (SCDCA), which is composed of multiple layers of convolution followed by a batch normalization layer and the leaky rectified linear unit (Leaky ReLU) activation function. Inspired by the idea of residual learning, we use two types of skip connections in the network. One is identity mapping between convolution layers and the other is used to connect the input and output. Through these connections, the network learns the residual between the noisy and clean images instead of learning an ordinary transformation function. We empirically evaluate our algorithm on an open and challenging document images dataset. We also assess our restoring results using the optical character recognition (OCR) test. Experimental results have demonstrated the effectiveness and efficiency of our proposed algorithm by comparing with several state-of-the-art methods.

Person re-identification(ReID) is a task of associating persons that cross the non-overlapping camera views at different locations and times. It is a challenging task due to the large variations in person pose, background, luminance, occlusion, low resolution, etc. How to extracting a powerful features representation is the prime problem in ReID and is still unsolved. In this paper, we propose a cascade network architecture combined with a generative adversarial networks(GANs) and a convolutional neural network(CNN) to improve the performance of person re-identification. The GANs first generates the person body parts segmentation from the person image, and then inputs the segmentation label into the connected CNN together with the original person image. Finally obtain a discriminative and robust feature representation for ReID task. The body parts segmentation partitioning the person image into multiple segments, such as background, head, face, arms, lags, etc. The body parts segmentation information contains accurate borders and category attributes for body parts, which makes the our model more accurate compared to other predefined rigid parts alignment models. Experiments are conduced on the CUHK03, Market1501, DukeMTMC-ReID datasets and the results demonstrate that this approach outperforms several existing state-of-the-art methods.

Zero-shot learning (ZSL) aims to recognize unseen object classes without any training samples, which can be regarded as a form of transfer learning from seen classes to unseen ones. This is made possible by learning a projection between a feature space and a semantic space (e.g. attribute space). Key to ZSL is thus to learn a projection function that is robust against the often large domain gap between the seen and unseen classes. In this paper, we propose a novel ZSL model termed domain-invariant projection learning (DIPL). Our model has two novel components: (1) A domain-invariant feature self-reconstruction task is introduced to the seen/unseen class data, resulting in a simple linear formulation that casts ZSL into a min-min optimization problem. Solving the problem is non-trivial, and a novel iterative algorithm is formulated as the solver, with rigorous theoretic algorithm analysis provided. (2) To further align the two domains via the learned projection, shared semantic structure among seen and unseen classes is explored via forming superclasses in the semantic space. Extensive experiments show that our model outperforms the state-of-the-art alternatives by significant margins.

Next basket recommendation is a new type of recommendation, which recommends a set of items, or a basket, to the user. Purchase in basket is a common behavior of consumers. Recently, deep neural networks have been applied to model sequential transactions of baskets in next basket recommendation. However, current methods do not track the user's evolving appetite for items explicitly, and they ignore important item attributes such as product category. In this paper, we propose a novel Attribute-aware Neural Attentive Model (ANAM) to address these problems. ANAM adopts an attention mechanism to explicitly model user's evolving appetite for items, and utilizes a hierarchical architecture to incorporate the attribute information. In specific, ANAM utilizes a recurrent neural network to model the user's sequential behavior over time, and relays the user's appetite toward items and their attributes to next basket through attention weights shared across baskets on the two different hierarchies. Experiment results on two public datasets (ıe Ta-Feng and JingDong) demonstrate the effectiveness of our ANAM model for next basket recommendation.

With the revival of neural networks, many studies try to adapt powerful sequential neural models, ıe Recurrent Neural Networks (RNN), to sequential recommendation. RNN-based networks encode historical interaction records into a hidden state vector. Although the state vector is able to encode sequential dependency, it still has limited representation power in capturing complicated user preference. It is difficult to capture fine-grained user preference from the interaction sequence. Furthermore, the latent vector representation is usually hard to understand and explain. To address these issues, in this paper, we propose a novel knowledge enhanced sequential recommender. Our model integrates the RNN-based networks with Key-Value Memory Network (KV-MN). We further incorporate knowledge base (KB) information to enhance the semantic representation of KV-MN. RNN-based models are good at capturing sequential user preference, while knowledge-enhanced KV-MNs are good at capturing attribute-level user preference. By using a hybrid of RNNs and KV-MNs, it is expected to be endowed with both benefits from these two components. The sequential preference representation together with the attribute-level preference representation are combined as the final representation of user preference. With the incorporation of KB information, our model is also highly interpretable. To our knowledge, it is the first time that sequential recommender is integrated with external memories by leveraging large-scale KB information.

Personalized PageRank (PPR) is a classic metric that measures the relevance of graph nodes with respect to a source node. Given a graph G, a source node s, and a parameter k, a top-k PPR query returns a set of k nodes with the highest PPR values with respect to s. This type of queries serves as an important building block for numerous applications in web search and social networks, such as Twitter's Who-To-Follow recommendation service. Existing techniques for top-k PPR, however, suffer from two major deficiencies. First, they either incur prohibitive space and time overheads on large graphs, or fail to provide any guarantee on the precision of top-k results (i.e., the results returned might miss a number of actual top-k answers). Second, most of them require significant pre-computation on the input graph G, which renders them unsuitable for graphs with frequent updates (e.g., Twitter's social graph).

In this paper, we present a large-scale sparse learning (LSSL) approach to solve the challenging task of semantic segmentation of images with noisy tags. Different from the traditional strongly supervised methods that exploit pixel-level labels for semantic segmentation, we make use of much weaker supervision (i.e., noisy tags of images) and then formulate the task of semantic segmentation as a weakly supervised learning (WSL) problem from the view point of noise reduction of superpixel labels. By learning the data manifolds, we transform the WSL problem into an LSSL problem. Based on nonlinear approximation and dimension reduction techniques, a linear-time-complexity algorithm is developed to solve the LSSL problem efficiently. We further extend the LSSL approach to visual feature refinement for semantic segmentation. The experiments demonstrate that the proposed LSSL approach can achieve promising results in semantic segmentation of images with noisy tags.

This article studies the problem of learning effective representations for Location-Based Social Networks (LBSN), which is useful in many tasks such as location recommendation and link prediction. Existing network embedding methods mainly focus on capturing topology patterns reflected in social connections, while check-in sequences, the most important data type in LBSNs, are not directly modeled by these models. In this article, we propose a representation learning method for LBSNs called as JRLM++, which models check-in sequences together with social connections. To capture sequential relatedness, JRLM++ characterizes two levels of sequential contexts, namely fine-grained and coarse-grained contexts. We present a learning algorithm tailored to the hierarchical architecture of the proposed model. We conduct extensive experiments on two important applications using real-world datasets. The experimental results demonstrate the superiority of our model. The proposed model can generate effective representations for both users and locations in the same embedding space, which can be further utilized to improve multiple LBSN tasks.

Online reviews have become an important source of information for users before making an informed purchase decision. Early reviews of a product tend to have a high impact on the subsequent product sales. In this paper, we take the initiative to study the behavior characteristics of early reviewers through their posted reviews on two real-world large e-commerce platforms, i.e., Amazon and Yelp. In specific, we divide product lifetime into three consecutive stages, namelyearly,majority, andlaggards. A user who has posted a review in the early stage is considered as an early reviewer. We quantitatively characterize early reviewers based on their rating behaviors, the helpfulness scores received from others and the correlation of their reviews with product popularity. We have found that (1) an early reviewer tends to assign a higher average rating score; and (2) an early reviewer tends to post more helpful reviews. Our analysis of product reviews also indicates that early reviewers’ ratings and their received helpfulness scores are likely to influence product popularity. By viewing the review posting process as a multiplayer competition game, we propose a novel margin-based embedding model for early reviewer prediction. Extensive experiments on two different e-commerce datasets have shown that our proposed approach outperforms a number of competitive baselines.

Search result diversification aims to retrieve diverse results to satisfy as many different information needs as possible. Supervised methods have been proposed recently to learn ranking functions and they have been shown to produce superior results to unsupervised methods. However, these methods use implicit approaches based on the principle of Maximal Marginal Relevance (MMR). In this paper, we propose a learning framework for explicit result diversification where subtopics are explicitly modeled. Based on the information contained in the sequence of selected documents, we use the attention mechanism to capture the subtopics to be focused on while selecting the next document, which naturally fits our task of document selection for diversification. As a preliminary attempt, we employ recurrent neural networks and max pooling to instantiate the framework. We use both distributed representations and traditional relevance features to model documents in the implementation. The framework is flexible to model query intent in either a flat list or a hierarchy. Experimental results show that the proposed method significantly outperforms all the existing search result diversification approaches.

Search result diversification aims at returning diversified document lists to cover different user intents of a query. Existing diversity measures assume that the intents of a query are disjoint, and do not consider their relationships. In this paper, we introduce intent hierarchies to model the relationships between intents, and present four weighing schemes. Based on intent hierarchies, we propose several hierarchical measures that take into account the relationships between intents. We demonstrate the feasibility of hierarchical measures by using a new test collection based on TREC Web Track 2009-2013 diversity test collections and by using NTCIR-11 IMine test collection. Our main experimental findings are: (1) Hierarchical measures are more discriminative and intuitive than existing measures. In terms of intuitiveness, it is preferable for hierarchical measures to use the whole intent hierarchies than to use only the leaf nodes. (2) The types of intent hierarchies used affect the discriminative power and intuitiveness of hierarchical measures. We suggest the best type of intent hierarchies to be used according to whether the nonuniform weights are available. (3) To measure the benefits of the diversification algorithms which use automatically mined hierarchical intents, it is important to use hierarchical measures instead of existing measures.

Online Judge (OJ) systems have been widely used in many areas, including programming, mathematical problems solving, and job interviews. Unlike other online learning systems, such as Massive Open Online Course, most OJ systems are designed for self-directed learning without the intervention of teachers. Also, in most OJ systems, problems are simply listed in volumes and there is no clear organization of them by topics or difficulty levels. As such, problems in the same volume are mixed in terms of topics or difficulty levels. By analyzing large-scale users’ learning traces, we observe that there are two major learning modes (or patterns). Users either practice problems in a sequential manner from the same volume regardless of their topics or they attempt problems about the same topic, which may spread across multiple volumes. Our observation is consistent with the findings in classic educational psychology. Based on our observation, we propose a novel two-mode Markov topic model to automatically detect the topics of online problems by jointly characterizing the two learning modes. For further predicting the difficulty level of online problems, we propose a competition-based expertise model using the learned topic information. Extensive experiments on three large OJ datasets have demonstrated the effectiveness of our approach in three different tasks, including skill topic extraction, expertise competition prediction and problem recommendation.

We propose a novel approach, called FeaBoost, to image semantic segmentation with only image-level labels taken as weakly-supervised constraints. Our approach is motivated from two evidences: 1) each superpixel can be represented as a linear combination of basic components (e.g., predefined classes); 2) visually similar superpixels have high probability to share the same set of labels, i.e., they tend to have common combination of predefined classes. By taking these two evidences into consideration, semantic segmentation is formulated as joint feature and label refinement over superpixels. Furthermore, we develop an efficient FeaBoost algorithm to solve such optimization problem. Extensive experiments on the MSRC and LabelMe datasets demonstrate the superior performance of our FeaBoost approach in comparison with the state-of-the-art methods, especially when noisy labels are provided for semantic segmentation.

More and more users prefer to ask their technical questions online. For machines, understanding a question is nontrivial. Current approaches lack explicit background knowledge.In this paper, we introduce a novel technical question understanding approach to recommending probable solutions to users. First, a knowledge graph is constructed which contains abundant technical information, and an augmented knowledge graph is built on the basis of the knowledge graph, to link the knowledge graph and documents. Then we develop a light weight question driven mechanism to select candidate documents. To improve the online performance, we propose an index-based random walk to support the online search. We use comprehensive experiments to evaluate the effectiveness of our approach on a large scale of real-world query logs. Our system outperforms main-stream search engine and the state-of-art information retrieval methods. Meanwhile, extensive experiments confirm the efficiency of our index-based online search mechanism.

Recently, deep neural networks have been widely applied to recommender systems. A representative work is to utilize deep learning for modeling complex user-item interactions. However, similar to traditional latent factor models by factorizing user-item interactions, they tend to be ineffective to capture localized information. Localized information, such as neighborhood, is important to recommender systems in complementing the user-item interaction data. Based on this consideration, we propose a novel Neighborhood-based Neural Collaborative Filtering model (NNCF). To the best of our knowledge, it is the first time that the neighborhood information is integrated into the neural collaborative filtering methods. Extensive experiments on three real-world datasets demonstrate the effectiveness of our model for the implicit recommendation task.

We propose and investigate a novel query, the Collective Travel Planning (CTP) query, that finds the lowest-cost route connecting multiple query sources and a destination via at most k meeting points. This type of query is useful in organizing large events, and it can bring significant benefits to society and the environment: it can help optimize the allocation of transportation resources, reduce resource consumption, and enable smarter and greener transportation; and it can help reduce greenhouse-gas emissions and traffic congestion.

This paper investigates the problem of weakly-supervised semantic segmentation, where image-level labels are used as weak supervision. Inspired by the successful use of Convolutional Neural Networks (CNNs) for fully-supervised semantic segmentation, we choose to directly train the CNNs over the oversegmented regions of images for weakly-supervised semantic segmentation. Although there are a few studies on CNNs-based weakly-supervised semantic segmentation, they have rarely considered the noise issue, i.e., the initial weak labels (e.g., social tags) may be noisy. To cope with this issue, we thus propose graph-boosted CNNs (GB-CNNs) for weakly-supervised semantic segmentation. In our GB-CNNs, the graph-based model provides the initial supervision for training the CNNs, and then the outcomes of the CNNs are used to retrain the graph-based model. This training procedure is iteratively implemented to boost the results of semantic segmentation. Experimental results demonstrate that the proposed model outperforms the state-of-the-art weakly-supervised methods. More notably, the proposed model is shown to be more robust in the noisy setting for weakly-supervised semantic segmentation.

Search result diversification aims to retrieve diverse results to satisfy as many different information needs as possible. Supervised methods have been proposed recently to learn ranking functions and they have been shown to produce superior results to unsupervised methods. However, these methods use implicit approaches based on the principle of Maximal Marginal Relevance (MMR). In this paper, we propose a learning framework for explicit result diversification where subtopics are explicitly modeled. Based on the information contained in the sequence of selected documents, we use attention mechanism to capture the subtopics to be focused on while selecting the next document, which naturally fits our task of document selection for diversification. The framework is implemented using recurrent neural networks and max-pooling which combine distributed representations and traditional relevance features. Our experiments show that the proposed method significantly outperforms all the existing methods.

The entity set expansion problem is to expand a small set of seed entities to a more complete set of similar entities. It can be applied in applications such as web search, item recommendation and query expansion. Traditionally, people solve this problem by exploiting the co-occurrence of entities within web pages, where latent semantic correlation among seed entities cannot be revealed. We propose a novel approach to solve the problem using knowledge graphs, by considering the deficiency (e.g., incompleteness) of knowledge graphs. We design an effective ranking model based on the semantic features of seeds to retrieve the candidate entities. Extensive experiments on public datasets show that the proposed solution significantly outperforms the state-of-the-art techniques.

Due to the rapid technological development of various sensors, a huge volume of high spatial resolution (HSR) image data can now be acquired. How to efficiently recognize the scenes from such HSR image data has become a critical task. Conventional approaches to remote sensing scene classification only utilize information from HSR images. Therefore, they always need a large amount of labeled data and cannot recognize the images from an unseen scene class without any visual sample in the labeled data. To overcome this drawback, we propose a novel approach for recognizing images from unseen scene classes, i.e., zero-shot scene classification (ZSSC). In this approach, we first use the well-known natural language process model, word2vec, to map names of seen/unseen scene classes to semantic vectors. A semantic-directed graph is then constructed over the semantic vectors for describing the relationships between unseen classes and seen classes. To transfer knowledge from the images in seen classes to those in unseen classes, we make an initial label prediction on test images by an unsupervised domain adaptation model. With the semantic-directed graph and initial prediction, a label-propagation algorithm is then developed for ZSSC. By leveraging the visual similarity among images from the same scene class, a label refinement approach based on sparse learning is used to suppress the noise in the zero-shot classification results. Experimental results show that the proposed approach significantly outperforms the state-of-the-art approaches in ZSSC.

A query facet is a significant list of information nuggets that explains an underlying aspect of a query. Existing algorithms mine facets of a query by extracting frequent lists contained in top search results. The coverage of facets and facet items mined by these kind of methods might be limited, because only a small number of search results are used. In order to solve this problem, we propose mining query facets by using knowledge bases which contain high-quality structured data. Specifically, we first generate facets based on the properties of the entities which are contained in Freebase and correspond to the query. Second, we mine initial query facets from search results, then expanding them by finding similar entities from Freebase. Experimental results show that our proposed method can significantly improve the coverage of facet items over the state-of-the-art algorithms.

With the increasing availability of moving-object tracking data, trajectory search is increasingly important. We propose and investigate a novel query type named trajectory search by regions of interest (TSR query). Given an argument set of trajectories, a TSR query takes a set of regions of interest as a parameter and returns the trajectory in the argument set with the highest spatial-density correlation to the query regions. This type of query is useful in many popular applications such as trip planning and recommendation, and location based services in general. TSR query processing faces three challenges: how to model the spatial-density correlation between query regions and data trajectories, how to effectively prune the search space, and how to effectively schedule multiple so-called query sources. To tackle these challenges, a series of new metrics are defined to model spatial-density correlations. An efficient trajectory search algorithm is developed that exploits upper and lower bounds to prune the search space and that adopts a query-source selection strategy, as well as integrates a heuristic search strategy based on priority ranking to schedule multiple query sources. The performance of TSR query processing is studied in extensive experiments based on real and synthetic spatial data.

Knowledge bases have shown their effectiveness in facilitating services like Web search and question-answering. Nevertheless, it remains challenging for ordinary users to fully understand the structure of a knowledge base and to issue structural queries. In many cases, users may have a natural language question and also know some popular (but not all) entities as sample answers. In this paper, we study the Reverse top-k Neighborhood Pattern Query problem, with the aim of discovering structural queries of the question based on: (i) the structure of the knowledge base, and (ii) the sample answers of the question. The proposed solution contains two phases: filter and refine. In the filter phase, a search space of candidate queries is systematically explored. The invalid queries whose result sets do not fully cover the sample answers are filtered out. In the refine phase, all surviving queries are verified to ensure that they are sufficiently relevant to the sample answers, with the assumption that the sample answers are more well-known or popular than other entities in the results of relevant queries. Several optimization techniques are proposed to accelerate the refine phrase. For evaluation, we conduct extensive experiments using the DBpedia knowledge base and a set of real-life questions. Empirical results show that our algorithm is able to provide a small set of possible queries, which contains the query matching the user question in natural language.

The prediction of periodical time-series remains challenging due to various types of scaling, misalignments and distortion effects. Here, we propose a novel model called Temporal embedding-enhanced convolutional neural Network (TeNet) to learn repeatedly-occurring-yet-hidden structural elements in periodical time-series, called abstract snippet detectors, to predict future changes. Our model effectively learns a new feature space for a time-series dataset. In the new feature space, distorted time-series that have implicit similarity but substantial differences in value and sequence to regular patterns are re-aligned to the regular patterns in the dataset, and subsequently contribute to a robust prediction mode. The model is robust to various types of distortions and misalignments and demonstrates strong prediction power for periodical time-series. We conduct extensive experiments and discover that the proposed model shows significant and consistent advantages over existing methods on a variety of data modalities ranging from human mobility to household power consumption records, when evaluated under four metrics. The model is also robust to various factors such as number of samples, variance of data, numerical ranges of data etc. The experiments verify that the intuition behind the model can be generalized to multiple data types and applications and promises significant improvement in prediction performance across the datasets studied.

Search result diversification aims at returning diversified document lists to cover different user intents for ambiguous or broad queries. Existing diversity measures assume that user intents are independent or exclusive, and do not consider the relationships among the intents. In this paper, we introduce intent hierarchies to model the relationships among intents. Based on intent hierarchies, we propose several hierarchical measures that can consider the relationships among intents. We demonstrate the feasibility of hierarchical measures by using a new test collection based on TREC Web Track 2009-2013 diversity test collections. Our main experimental findings are: (1) Hierarchical measures are generally more discriminative and intuitive than existing measures using flat lists of intents; (2) When the queries have multilayer intent hierarchies, hierarchical measures are less correlated to existing measures, but can get more improvement in discriminative power; (3) Hierarchical measures are more intuitive in terms of diversity or relevance. The hierarchical measures using the whole intent hierarchies are more intuitive than only using the leaf nodes in terms of diversity and relevance.

Large-scale matrix computation becomes essential for many data data applications, and hence the problem of sketching matrix with small space and high precision has received extensive study for the past few years. This problem is often considered in the row-update streaming model, where the data set is a matrix A -- Rn x d, and the processor receives a row (1 x d) of A at each timestamp. The goal is to maintain a smaller matrix (termed approximation matrix, or simply approximation) B -- Rl x d as an approximation to A, such that the covariance error |AT A - BTB| is small and l ll n.