Large language models (LLMs) have achieved remarkable success in a wide range of tasks. However, their reasoning capabilities, particularly in complex domains like mathematics, remain a significant challenge. Value-based process verifiers, which estimate the probability of a partial reasoning chain leading to a correct solution, are a promising approach for improving reasoning. Nevertheless, their effectiveness is often hindered by estimation error in their training annotations, a consequence of the limited number of Monte Carlo (MC) samples feasible due to the high cost of LLM inference. In this paper, we identify that the estimation error primarily arises from high variance rather than bias, and the MC estimator is a Minimum Variance Unbiased Estimator (MVUE). To address the problem, we propose the Compound Monte Carlo Sampling (ComMCS) method, which constructs an unbiased estimator by linearly combining the MC estimators from the current and subsequent steps. Theoretically, we show that our method leads to a predictable reduction in variance, while maintaining an unbiased estimation without additional LLM inference cost. We also perform empirical experiments on the MATH-500 and GSM8K benchmarks to demonstrate the effectiveness of our method. Notably, ComMCS outperforms regression-based optimization method by 2.8 points, the non-variance-reduced baseline by 2.2 points on MATH-500 on Best-of-32 sampling experiment.

In the literature, existing human-centric emotional motion generation methods primarily focus on boosting performance within a single scale-fixed dataset, largely neglecting the flexible and scale-increasing motion scenarios (e.g., sports, dance), whereas effectively learning these newly emerging scenarios can significantly enhance the model’s real-world generalization ability. Inspired by this, this paper proposes a new LLM-Centric Lifelong Empathic Motion Generation (L2-EMG) task, which aims to equip LLMs with the capability to continually acquire emotional motion generation knowledge across different unseen scenarios, potentially contributing to building a closed-loop and self-evolving embodied agent equipped with both empathy and intelligence. Further, this paper poses two key challenges in the L2-EMG task, i.e., the emotion decoupling challenge and the scenario adapting challenge. To this end, this paper proposes an Emotion-Transferable and Scenario-Adapted Mixture of Experts (ES-MoE) approach which designs a causal-guided emotion decoupling block and a scenario-adapted expert constructing block to address the two challenges, respectively. Especially, this paper constructs multiple L2-EMG datasets to validate the effectiveness of the ES-MoE approach. Extensive evaluations show that ES-MoE outperforms advanced baselines.

News image captioning aims to produce journalistically informative descriptions by combining visual content with contextual cues from associated articles. Despite recent advances, existing methods struggle with three key challenges: (1) incomplete information coverage, (2) weak cross-modal alignment, and (3) suboptimal visual-entity grounding. To address these issues, we introduce MERGE, the first Multimodal Entity-aware Retrieval-augmented GEneration framework for news image captioning. MERGE constructs an entity-centric multimodal knowledge base (EMKB) that integrates textual, visual, and structured knowledge, enabling enriched background retrieval. It improves cross-modal alignment through a multistage hypothesis-caption strategy and enhances visual-entity matching via dynamic retrieval guided by image content. Extensive experiments on GoodNews and NYTimes800k show that MERGE significantly outperforms state-of-the-art baselines, with CIDEr gains of +6.84 and +1.16 in caption quality, and F1-score improvements of +4.14 and +2.64 in named entity recognition. Notably, MERGE also generalizes well to the unseen Visual News dataset, achieving +20.17 in CIDEr and +6.22 in F1-score, demonstrating strong robustness and domain adaptability.

Large Language Models (LLMs) face severe challenges in long-context processing, including quadratic computational costs, information forgetting, and the context fragmentation inherent in Retrieval-Augmented Generation (RAG).We introduce LycheeMemory, a cognitively inspired framework that enables efficient long-context inference via chunk-wise compression and selective memory recall, rather than processing all raw tokens.LycheeMemory segments the input into chunks and encodes each into compressed KV-cache-style representations using a Compressor.A Gate then dynamically selects relevant memory blocks, which a Reasoner iteratively processes with an evolving working memory to solve downstream tasks.The Compressor and Reasoner are jointly optimized via end-to-end reinforcement learning, while the Gate is trained separately as a classifier.Experimental results demonstrate that LycheeMemory achieves competitive accuracy (up to 82% in ablation variants) on multi-hop reasoning benchmarks (e.g., RULER-HQA), successfully extrapolates context length from 7K to 1.75M, and provides a favorable accuracy-efficiency trade-off against strong long-context baselines.Notably, compared to MemAgent, LycheeMemory achieves an average 2× reduction in peak GPU memory usage and a 6× speedup during inference.

Existing long-document question answering systems typically process texts as flat sequences or use heuristic chunking, which overlook the discourse structures that naturally guide human comprehension.We present a discourse-aware hierarchical framework that leverages rhetorical structure theory (RST) for long document question answering.Our approach converts discourse trees into sentence-level representations and employs LLM-enhanced node representations to bridge structural and semantic information.The framework involves three key innovations: language-universal discourse parsing for lengthy documents, LLM-based enhancement of discourse relation nodes, and structure-guided hierarchical retrieval.Extensive experiments on four datasets demonstrate consistent improvements over existing approaches through the incorporation of discourse structure, across multiple genres and languages.Moreover, the proposed framework exhibits strong robustness across diverse document types and linguistic settings.

Linxuan Du, Guangquan Xue, Xiaobo Liang, Qipeng Huang, Yuyang Ding, Xinyu Shi, Zhang Yijun, Ji Qi, Wenpeng Zhu, Juntao Li, Min Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Weiyang Huang, Xuefeng Bai, Kehai Chen, Xinyang Chen, Yibin Chen, Weili Guan, Min Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Large Language Models (LLMs) enhance their problem-solving capability by utilizing external tools.However, in open-world scenarios with massive and evolving tool repositories, existing methods relying on static embedding retrieval or parameter memorization of tools struggle to align user intent with tool semantics or generalize to unseen tools, respectively, leading to suboptimal accuracy of open-world tool retrieval and execution.To address these, we present ToolOmni, a unified agentic framework that enables LLMs for open-world tool use by proactive retrieval and grounded execution within a reasoning loop.First, we construct a cold-start multi-turn interaction dataset to instill foundational agentic capabilities via Supervised Fine-Tuning (SFT).Then, we introduce open-world tool learning based on a Decoupled Multi-Objective GRPO algorithm, which simultaneously optimizes LLMs for both tool retrieval accuracy and execution efficacy in online environments.Extensive experiments demonstrate that ToolOmni achieves state-ofthe-art performance both in retrieval and execution, surpassing strong baselines by a significant margin of +10.8% in end-to-end execution success rate, while exhibiting exceptional robustness and generalization capabilities.

Rongqing Jiang, Xuebo Liu, Shengxin Liu, Yutong Wang, Min Zhang, Shimin Tao, Daimeng Wei, Min Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Bobo Li, Wu Rui, Zibo Ji, Meishan Zhang, Hao Fei, Min Zhang, Mong-Li Lee, Wynne Hsu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Niu Lian, Yuting Wang, Hanshu Yao, Jinpeng Wang, Bin Chen, Yaowei Wang, Min Zhang, Shu-Tao Xia. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Xiaobo Liang, Wanfu Wang, Qipeng Huang, Yuyang Ding, Zecheng Tang, Yixin Ji, Qianben Chen, Zhe Zhao, Kehai Chen, Juntao Li, Min Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Gewen Liang, Mufan Xu, Kehai Chen, Wei Wang, Yuwei Wang, Muyun Yang, Tiejun Zhao, Min Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Zhenyu Liu, Yunxin li, Xuanyu Zhang, Qixun Teng, Shenyuan Jiang, Xinyu Chen, Haoyuan Shi, Haolan Chen, Fanbo Meng, Mingjun Zhao, Yu Xu, Yancheng He, Baotian Hu, Haizhou Li, Min Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Zhenyu Liu, Xuanyu Zhang, Yunxin li, Qixun Teng, Shenyuan Jiang, Haolan Chen, Mingjun Zhao, Fanbo Meng, Yu Xu, Yancheng He, Baotian Hu, Haizhou Li, Min Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Yi Su, Dian Yu, Linfeng Song, Juntao Li, Haitao Mi, Zhaopeng Tu, Min Zhang, Dong Yu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Zechen Sun, Yuyang Sun, Zecheng Tang, Juntao Li, Wenpeng Hu, Wenliang Chen, Zhunchen Luo, Guotong Geng, Min Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Multimodal Stance Detection (MSD) is a crucial task for understanding public opinion on social media.Existing methods predominantly operate by learning to fuse modalities.They lack an explicit reasoning process to discern how inter-modal dynamics, such as irony or conflict, collectively shape the user's final stance, leading to frequent misjudgments.To address this, we advocate for a paradigm shift from learning to fuse to learning to reason.We introduce MIND, a Meta-cognitive Intuitivereflective Network for Dual-reasoning.Inspired by the dual-process theory of human cognition, MIND operationalizes a self-improving loop.It first generates a rapid, intuitive hypothesis by querying evolving Modality and Semantic Experience Pools.Subsequently, a meta-cognitive reflective stage uses Modality-CoT and Semantic-CoT to scrutinize this initial judgment, distill superior adaptive strategies, and evolve the experience pools themselves.These dual experience structures are continuously refined during training and recalled at inference to guide robust and context-aware stance decisions.Extensive experiments on the MMSD benchmark demonstrate that our MIND significantly outperforms most baseline models and exhibits strong generalization.

Yue Wang, Ruotian Ma, Xingyu Chen, Zhengliang Shi, Morunliu Yang, Wanshun Chen, Huang Liu, Jiadi Yao, Xin He, Qu Yang, Qingxuan Jiang, Fanghua Ye, Juntao Li, Zhaopeng Tu, Xiaolong Li, Liefeng Bo, Min Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Large reasoning models (LRMs) have achieved remarkable performance in complex reasoning tasks, driven by their powerful inference-time scaling capability.However, LRMs often suffer from overthinking, which results in substantial computational redundancy and significantly reduces efficiency.Early-exit methods aim to mitigate this issue by terminating reasoning once sufficient evidence has been generated, yet existing approaches mostly rely on handcrafted or empirical indicators that are unreliable and impractical.In this work, we introduce Dynamic Thought Sufficiency in Reasoning (DTSR), a novel framework for efficient reasoning that enables the model to dynamically assess the sufficiency of its chain-of-thought (CoT) and determine the optimal point for early exit.Inspired by human metacognition, DTSR operates in two stages: (1) Reflection Signal Monitoring, which identifies reflection signals as potential cues for early exit, and (2) Thought Sufficiency Check, which evaluates whether the current CoT is sufficient to derive the final answer.Experimental results on the Qwen3 models show that DTSR reduces reasoning length by 28.9%-34.9%with minimal performance loss, effectively mitigating overthinking.We further discuss overconfidence in LRMs and self-evaluation paradigms, providing valuable insights for early-exit reasoning.

Chenyuan Zhang, Qiguang Chen, Xie Chen, Zhuotao Tian, Bowen Xing, Meishan Zhang, Libo Qin, Baotian Hu, Min Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Large language model driven multi-agent social simulation frameworks enable realistic modeling of complex societal dynamics but incur substantial computational overhead due to dense agent participation and extensive interaction costs. To address this limitation, we propose SocialDropout, a reinforcement learning–based agent selection strategy within the AgentSociety framework, inspired by the AgentDropout paradigm, which dynamically identifies and samples informative agent subsets for each simulation round. Each agent is assigned an adaptive importance weight optimized to jointly minimize agent sparsity and computational cost measured by LLM calls, token consumption, and execution time—while preserving social interaction intensity within the environment. Extensive performance evaluation demonstrates that the proposed method significantly improves simulation efficiency and scalability. Moreover, ablation studies verify that high-level behavioral realism and outcome consistency are largely maintained despite substantial agent reduction. Our approach offers a practical and general optimization mechanism for large-scale LLM-based multi-agent social simulations under constrained computational budgets.

Contextual meme understanding decodes implicit meaning from ambiguous visual metaphors and social context, which is crucial for online communication analysis. However, existing linear reasoning paradigms reductively cast this process as deterministic decoding, overlooking the fundamental reality that meme interpretation necessitates the dynamic alignment of multimodal cues. Lacking mechanisms to measure and correct misalignment, these static models inevitably allow initial perceptual failures to cascade into irreversible hallucinations. To address this, we propose Semantic Energy Entropy Descent (SEED), a framework that reformulates contextual meme understanding as an energy minimization problem within a discrete semantic space. Specifically, SEED constructs a Convergent Heterogeneous Thought Tree (CHTT) as the optimization workspace, where an Evaluator Agent quantifies the semantic inconsistency of hypotheses via a Semantic Energy Mechanism. By calculating a Discrete Semantic Gradient as structured feedback, the framework activates a Supplier Agent to retrieve knowledge and context evidence and a Reasoner Agent to refine explanations, thereby orchestrating Semantic Descent Dynamics that iteratively drive the reasoning trajectory to converge on a stable, low-entropy interpretation. Experimental results show that SEED consistently outperforms strong baselines on both classification and generation, reducing logical hallucinations and improving cultural grounding.

Text-to-image person retrieval (TIPR) aims to identify the target person using textual descriptions, facing challenge in modality heterogeneity. Prior works have attempted to address it by developing cross-modal global or local alignment strategies. However, global methods typically overlook fine-grained cross-modal differences, whereas local methods require prior information to explore explicit part alignments. Additionally, current methods are English-centric, restricting their application in multilingual contexts. To alleviate these issues, we pioneer a multilingual TIPR task by developing a multilingual TIPR benchmark, for which we leverage large language models for initial translations and refine them by integrating domain-specific knowledge. Correspondingly, we propose Bi-IRRA: a Bidirectional Implicit Relation Reasoning and Aligning framework to learn alignment across languages and modalities. Within Bi-IRRA, a bidirectional implicit relation reasoning module enables bidirectional prediction of masked image and text, implicitly enhancing the modeling of local relations across languages and modalities, a multi-dimensional global alignment module is integrated to bridge the modality heterogeneity. The proposed method achieves new state-of-the-art results on all multilingual TIPR datasets.

Multi-teacher knowledge distillation transfers knowledge from multiple large teacher models to a small student model and has performed well on many downstream tasks. However, when distilling knowledge from multiple teachers, it always suffers from the severe problems of being time-consuming and storage-extensive for multiple teacher models training and inference. We present MoE-KD, a simple but effective framework that produces supervision for training the student model from one single teacher model, which fixes the above problems and improves effectiveness. In the proposed MoE-KD, multiple trainable prompts are used to extract different views of samples from a single pre-trained language model and only a few parameters (prompts) need to be trained and stored. To guarantee the generated supervision signals with increased robustness and correctness, we introduce an uncertainty-based mechanism and a selector module, which routes the input instance to its corresponding teacher. We have also extended MoE KD to lifelong learning scenarios, proposing a lightweight solution for catastrophic forgetting. We conduct experiments on traditional KD scenarios and lifelong learning scenarios. MoE-KD yields improvements up to 1.1% and 140% in accuracy and efficiency in knowledge distillation and 2.8% improvements on average in lifelong learning, compared with the strong baseline methods.

Large language models (LLMs) have created a new paradigm for natural language processing. Despite their advancement, LLM-based methods still lag behind traditional approaches in document-level relation extraction (DocRE), a critical task for understanding complex entity relations within long context. This paper investigates the causes of this performance gap, identifying the dispersion of attention by LLMs due to entity pairs without relations as a key factor. We then introduce a novel classifier-LLM approach to DocRE. Particularly, the proposed approach begins with a classifier designed to select entity pair candidates that exhibit potential relations and then feed them to LLM for final relation classification. This method ensures that the LLM's attention is directed at relation-expressing entity pairs instead of those without relations during inference. Experiments on DocRE benchmarks reveal that our method significantly outperforms recent LLM-based DocRE models and narrows the performance gap with state-of-the-art BERT-based models.

Recent advancements in multimodal large language models (MLLMs) have achieved significant multimodal generation capabilities, akin to GPT-4. These models predominantly map visual information into language representation space, leveraging the vast knowledge and powerful text generation abilities of LLMs to produce multimodal instruction-following responses. We could term this method as LLMs for Vision because of its employing LLMs for visual understanding and reasoning, yet observe that these MLLMs neglect the potential of harnessing visual knowledge to enhance the overall capabilities of LLMs, which could be regarded as Vision Enhancing LLMs. In this paper, we propose an approach called MKS2, aimed at enhancing LLMs through empowering Multimodal Knowledge Storage and Sharing in LLMs. Specifically, we introduce Modular Visual Memory (MVM), a component integrated into the internal blocks of LLMs, designed to store open-world visual information efficiently. Additionally, we present a soft Mixture of Multimodal Experts (MoMEs) architecture in LLMs to invoke multimodal knowledge collaboration during text generation. Our comprehensive experiments demonstrate that MKS2 substantially augments the reasoning capabilities of LLMs in contexts necessitating physical or commonsense knowledge. It also delivers competitive results on image-text understanding multimodal benchmarks. The codes will be available at: https://github.com/HITsz-TMG/MKS2-Multimodal-Knowledge-Storage-and-Sharing.

A neural Multimodal Machine Translation (MMT) system utilizes multimodal information, particularly images, to enhance traditional text-only models and achieve superior performance. However, the effectiveness of MMT heavily depends on the availability of extensive collections of bilingual parallel sentence pairs and manually annotated images, which poses a challenge due to the scarcity of such pairs. To address this issue, we propose incorporating the Multimodal Knowledge Graph (MMKG) for data augmentation in MMT. By utilizing MMKG as an additional source of knowledge, we can overcome the limitations of existing sentence-image pairings. This allows us to expand the original parallel corpus and generate corresponding images, creating new synthetic data pairs that facilitate effective data augmentation. Experiments conducted on two translation datasets, Multi30k and IKEA, demonstrate that the proposed MMKG enhancement method significantly improves performance across multiple baseline methods, ultimately outperforming all baseline approaches. Additionally, experiments under low-resource conditions reveal that our method achieves exceptional enhancement effects in low-resource corpora, surpassing other data augmentation baseline methods. These results indicate the efficacy and potential of the proposed method for enhancing the performance of multimodal models across diverse datasets.

Large language models (LLMs) are typically adapted for context-aware machine translation (MT) by combining both the source sentence and its surrounding sentences into a single input. This unified input is then processed in one go, with the model producing the target translation step by step. However, this method treats the intrasentence and intersentence contexts similarly, even though they play distinct roles. In this study, we present a novel strategy called multiphase prompt tuning (MPT) to address this issue by enabling LLMs to treat these two context types differently. MPT divides the context-aware translation task into three phases: encoding the intersentence context, encoding the source sentence, and the final decoding phase. Each phase incorporates distinct continuous prompts that help the model focus on the appropriate task for each type of context. We also introduce a multitask fine-tuning approach to emphasize the distinction between intersentence and intrasentence contexts and enhance intersentence dependencies. This includes two auxiliary tasks: context-agnostic translation and cross-lingual next sentence generation, which help extract additional information and improve the model's handling of discourse-related challenges.

Despite increasing research attention to text ranking, most studies focus on monolingual scenarios, with a particular emphasis on English-language contexts. This narrow focus limits the applicability of ranking models in cross-lingual contexts, such as ranking Chinese documents based on English queries. Recent advances in large language models (LLMs) have significantly reduced inter-language barriers through pre-training on extensive multilingual corpora, thus facilitating the study of multilingual text ranking (MTR). In this work, we explore the potential of LLMs in MTR tasks. Specifically, we first introduce an MTR benchmark encompassing both monolingual and cross-lingual scenarios. Then, we propose a two-stage training pipeline to alleviate the misalignment between LLMs and text ranking. Lastly, we adapt this training pipeline to multilingual scenarios from the perspective of training data and methods. Our experiments on the MTR benchmark demonstrate that the proposed multilingual two-stage training pipeline significantly improves LLM ranking performance in both monolingual and cross-lingual scenarios, particularly in out-domain settings. We complement these findings with a thorough analysis to deepen the understanding of our approach.

Large Language Models (LLMs) need to adapt to the continuous changes in data, tasks, and user preferences. Due to their massive size and the high costs associated with training, LLMs are not suitable for frequent retraining. However, updates are necessary to keep them in sync with rapidly evolving human knowledge. To address these challenges, this paper proposes the Compression Memory Training (CMT) method, an efficient and effective online adaptation framework for LLMs that features robust knowledge retention capabilities. Inspired by human memory mechanisms, CMT compresses and extracts information from new documents to be stored in a memory bank. When answering to queries related to these new documents, the model aggregates these document memories from the memory bank to better answer user questions. The parameters of the LLM itself do not change during training and inference, reducing the risk of catastrophic forgetting. To enhance the encoding, retrieval, and aggregation of memory, we further propose three new general and flexible techniques, including memory-aware objective, self-matching and top-k aggregation. Extensive experiments conducted on three continual learning datasets (i.e., StreamingQA, SQuAD and ArchivalQA) demonstrate that the proposed method improves model adaptability and robustness across multiple base LLMs (e.g., +4.07 EM & +4.19 F1 in StreamingQA with Llama-2-7b).

Multi-hop question answering (MHQA) poses a significant challenge for large language models (LLMs) due to the extensive knowledge demands involved. Knowledge editing, which aims to precisely modify the LLMs to incorporate specific knowledge without negatively impacting other unrelated knowledge, offers a potential solution for addressing MHQA challenges with LLMs. However, current solutions struggle to effectively resolve issues of knowledge conflicts. Most parameter-preserving editing methods are hindered by inaccurate retrieval and overlook secondary editing issues, which can introduce noise into the reasoning process of LLMs. In this paper, we introduce KEDKG, a novel knowledge editing method that leverages a dynamic knowledge graph for MHQA, designed to ensure the reliability of answers. KEDKG involves two primary steps: dynamic knowledge graph construction and knowledge graph augmented generation. Initially, KEDKG autonomously constructs a dynamic knowledge graph to store revised information while resolving potential knowledge conflicts. Subsequently, it employs a fine-grained retrieval strategy coupled with an entity and relation detector to enhance the accuracy of graph retrieval for LLM generation. Experimental results on benchmarks show that KEDKG surpasses previous state-of-the-art models, delivering more accurate and reliable answers in environments with dynamic information.

Despite being empowered with alignment mechanisms, large language models (LLMs) are increasingly vulnerable to emerging jailbreak attacks that can compromise their alignment mechanisms. This vulnerability poses significant risks to real-world applications. Existing work faces challenges in both training efficiency and generalization capabilities (i.e., Reinforcement Learning from Human Feedback and Red-Teaming). Developing effective strategies to enable LLMs to resist continuously evolving jailbreak attempts represents a significant challenge. To address this challenge, we propose a novel defensive paradigm called GuidelineLLM, which assists LLMs in recognizing queries that may have harmful content. Before LLMs respond to a query, GuidelineLLM first identifies potential risks associated with the query, summarizes these risks into guideline suggestions, and then feeds these guidelines to the responding LLMs. Importantly, our approach eliminates the necessity for additional safety fine-tuning of the LLMs themselves; only the GuidelineLLM requires fine-tuning. This characteristic enhances the general applicability of GuidelineLLM across various LLMs. Experimental results demonstrate that GuidelineLLM can significantly reduce the attack success rate (ASR) against LLM (an average reduction of 34.17% ASR) while maintaining the usefulness of LLM in handling benign queries.

Chinese spelling correction (CSC) is a crucial task that aims to correct character errors in Chinese text.While conventional CSC focuses on character substitution errors caused by mistyping, two other common types of character errors, missing and redundant characters, have received less attention.These errors are often excluded from CSC datasets during the annotation process or ignored during evaluation, even when they have been annotated.This issue limits the practicality of the CSC task.To address this issue, we introduce the task of General Chinese Character Error Correction (C2EC), which focuses on all three types of character errors.We construct a high-quality C2EC benchmark by combining and manually verifying data from CCTC and Lemon datasets.We extend the training-free prompt-free CSC method to C2EC by using Levenshtein distance for handling length changes and leveraging an additional prompt-based large language model (LLM) to improve performance.Experiments show that our method enables a 14B-parameter LLM to be on par with models nearly 50 times larger on both conventional CSC and C2EC tasks, without any fine-tuning.

Guodong Du, Zitao Fang, Jing Li, Junlin Li, Runhua Jiang, Shuyang Yu, Yifei Guo, Yangneng Chen, Sim Kuan Goh, Ho-Kin Tang, Daojing He, Honghai Liu, Min Zhang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

The impressive capabilities of Large Language Models (LLMs) come at the cost of substantial computational resources during deployment. While KV Cache can significantly reduce recomputation during inference, it also introduces additional memory overhead. KV Cache quantization presents a promising solution, striking a good balance between memory usage and accuracy. Previous research has shown that the Keys are distributed by channel, while the Values are distributed by token. Consequently, the common practice is to apply channel-wise quantization to the Keys and token-wise quantization to the Values. However, our further investigation reveals that a small subset of unusual tokens exhibit unique characteristics that deviate from this pattern, which can substantially impact quantization accuracy. To address this, we develop a simple yet effective method to identify these tokens accurately during the decoding process and exclude them from quantization as outlier tokens, significantly improving overall accuracy. Extensive experiments show that our method achieves significant accuracy improvements under 2-bit quantization and can deliver a 6.4 times reduction in memory usage and a 2.3 times increase in throughput.

Attributed text generation aims to enhance the reliability of content generated from large language models by providing citations for each claim, which thereby enables users to easily verify the correctness of the responses.However, the scarcity of high-quality training samples presents a significant challenge in aligning large language models to generate texts with citations, revealing considerable room for improvement in existing attribution systems.Besides, existing approaches of aligning large language models to follow user instructions can lead to an undue emphasis on irrelevant documents, which in turn reduces the quality of responses.To address the above problems, we propose Automatic Alignment Framework for Attributed Text Generation (A 3 ), a novel framework designed to automatically generate highquality attributed query-response pairs for both supervised fine-tuning and preference optimization stages without human annotation.With the help of A 3 , Mistral-7B can achieve a citation recall of 84.4 and a precision of 87.0 precision on ASQA, which notably surpasses GPT-4's citation recall of 73.0 and precision of 76.5. 1

Xinyu Chen, Yunxin Li, Haoyuan Shi, Baotian Hu, Wenhan Luo, Yaowei Wang, Min Zhang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Andong Chen, Yuchen Song, Kehai Chen, Xuefeng Bai, Muyun Yang, Liqiang Nie, Jie Liu, Tiejun Zhao, Min Zhang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Reinforcement Learning (RL) algorithms for safety alignment of Large Language Models (LLMs), such as Direct Preference Optimization (DPO), encounter the challenge of distribution shift.Current approaches typically address this issue through online sampling from the target policy, which requires significant computational resources.In this paper, we hypothesize that during off-policy training, while the ranking order of output generated by policy changes, their overall distribution remains relatively stable.This stability allows the conversion of the sampling process from the target policy into a computationally efficient reranking of preference data.Building on this hypothesis, we propose a new framework that leverages the model's intrinsic safety judgment capability to extract reward signals, which are then used to calculate label confidence for preference reordering.Extensive experiments and theoretical analysis demonstrate that the proposed method effectively addresses the distribution shift issue, remarkably enhancing the safety performance while avoiding about 300x computational overheads.

Large language models (LLMs) deliver impressive results but face challenges from increasing model sizes and computational costs.Structured pruning reduces model size and speeds up inference but often causes uneven degradation across domains, leading to biased performance.To address this, we propose DR-Pruning, a method that dynamically adjusts the data distribution during training to restore balanced performance across heterogeneous and multi-tasking data.Experiments in monolingual and multilingual settings show that DR-Pruning surpasses similarly sized models in both pruning and continued pretraining over perplexity, downstream tasks, and instruction tuning.Further analysis demonstrates the robustness of DRPruning towards various domains and distribution shifts.Furthermore, DRPruning can determine optimal reference losses and data ratios automatically, suggesting potential for broader applications.

Improving the mathematical reasoning capabilities of Large Language Models (LLMs) is critical for advancing artificial intelligence. However, access to extensive, diverse, and high-quality reasoning datasets remains a significant challenge, particularly for the open-source community. In this paper, we propose ScaleQuest, a novel, scalable, and cost-effective data synthesis method that enables the generation of large-scale mathematical reasoning datasets using lightweight 7B-scale models. ScaleQuest introduces a two-stage question-tuning process comprising Question Fine-Tuning (QFT) and Question Preference Optimization (QPO) to unlock the question generation capabilities of problem-solving models. By generating diverse questions from scratch – without relying on powerful proprietary models or seed data – we produce a dataset of 1 million problem-solution pairs. Our experiments demonstrate that models trained on our data outperform existing open-source datasets in both in-domain and out-of-domain evaluations. Furthermore, our approach shows continued performance improvement as the volume of training data increases, highlighting its potential for ongoing data scaling. The extensive improvements observed in code reasoning tasks demonstrate the generalization capabilities of our proposed method. Our work provides the open-source community with a practical solution to enhance the mathematical reasoning abilities of LLMs.

The proliferation of jailbreak attacks against large language models (LLMs) highlights the need for robust security measures.However, in multi-round dialogues, malicious intentions may be hidden in interactions, leading LLMs to be more prone to produce harmful responses.In this paper, we propose the Multi-Turn Safety Alignment (MTSA) framework, to address the challenge of securing LLMs in multi-round interactions.It consists of two stages: In the thought-guided attack learning stage, the redteam model learns about thought-guided multiround jailbreak attacks to generate adversarial prompts.In the adversarial iterative optimization stage, the red-team model and the target model continuously improve their respective capabilities in interaction.Furthermore, we introduce a multi-turn reinforcement learning algorithm based on future rewards to enhance the robustness of safety alignment.Experimental results show that the red-team model exhibits state-of-the-art attack capabilities, while the target model significantly improves its performance on safety benchmarks.

Cross-domain constituency parsing is still an unsolved challenge in computational linguistics since the available multi-domain constituency treebank is limited.We investigate automatic treebank generation by large language models (LLMs) in this paper.The performance of LLMs on constituency parsing is poor, therefore we propose a novel treebank generation method, LLM back generation, which is similar to the reverse process of constituency parsing.LLM back generation takes the incomplete cross-domain constituency tree with only domain keyword leaf nodes as input and fills the missing words to generate the cross-domain constituency treebank.Besides, we also introduce a span-level contrastive learning pretraining strategy to make full use of the LLM back generation treebank for cross-domain constituency parsing.We verify the effectiveness of our LLM back generation treebank coupled with contrastive learning pre-training on five target domains of MCTB.Experimental results show that our approach achieves state-of-theart performance on average results compared with various baselines.

Junlin Li, Guodong Du, Jing Li, Sim Kuan Goh, Wenya Wang, Yequan Wang, Fangming Liu, Ho-Kin Tang, Saleh Alharbi, Daojing He, Min Zhang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Generative Reward Models (GenRMs) leverage synthesized Chains of Thought (CoT) to reduce the need for massive labeled data, but this approach introduces risks of overoptimization due to the inability to guarantee the correctness of the CoTs.Identifying and optimizing unexpected behaviors within these synthesized CoT remains a challenge, as it heavily depends on precise annotations of intermediate behavior, similar to process supervision.In this work, we introduce a criteria-based preference tree for reward modeling, where each path in the tree represents a reasoning trajectory based on synthesized criteria.Crucially, each reasoning trajectory can be independently optimized through RL algorithm.These fine-grained process reward signals are derived from the inferencetime computations and predefined rules, eliminating the need for human supervision.In experiments, SyncPL 1 showed significant improvements over baselines on multiple human preference benchmarks.We further demonstrate that synthesized data can be learned using a long CoT format, analogous to an o1-like model, further enhancing performance while keeping stability and efficiency during training.

Ziheng Qiao, Houquan Zhou, Yumeng Liu, Zhenghua Li, Min Zhang, Bo Zhang, Chen Li, Ji Zhang, Fei Huang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Zesheng Shi, Yucheng Zhou, Jing Li, Yuxin Jin, Yu Li, Daojing He, Fangming Liu, Saleh Alharbi, Jun Yu, Min Zhang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Long-context models (LCMs) have witnessed remarkable advancements in recent years, facilitating real-world tasks like long-document QA.The success of LCMs is founded on the hypothesis that the model demonstrates strong fidelity, enabling it to respond based on the provided long context rather than relying solely on the intrinsic knowledge acquired during pre-training.Yet, in this paper, we find that open-sourced LCMs are not as faithful as expected.We introduce L-CiteEval, an out-of-the-box suite that can assess both generation quality and fidelity in long-context understanding tasks.It covers 11 tasks with context lengths ranging from 8K to 48K and a corresponding automatic evaluation pipeline.Evaluation of 11 cutting-edge closed-source and open-source LCMs indicates that, while there are minor differences in their generation, open-source models significantly lag behind closed-source counterparts in terms of fidelity.Furthermore, we analyze the benefits of citation generation for LCMs from both the perspective of explicit model output and the internal attention mechanism 1 .

Multi-agent systems (MAS) based on large language models (LLMs) have demonstrated significant potential in collaborative problemsolving.However, they still face substantial challenges of low communication efficiency and suboptimal task performance, making the careful design of the agents' communication topologies particularly important.Inspired by the management theory that roles in an efficient team are often dynamically adjusted, we propose AgentDropout, which identifies redundant agents and communication across different communication rounds by optimizing the adjacency matrices of the communication graphs and eliminates them to enhance both token efficiency and task performance.Compared to state-of-the-art methods, AgentDropout achieves an average reduction of 21.6% in prompt token consumption and 18.4% in completion token consumption, along with a performance improvement of 1.14 on the tasks.Furthermore, the extended experiments demonstrate that AgentDropout achieves notable domain transferability and structure robustness, revealing its reliability and effectiveness.We release our code at https://github. com/wangzx1219/AgentDropout.

Conditional image generation has gained significant attention for its ability to personalize content. However, the field faces challenges in developing task-agnostic, reliable, and explainable evaluation metrics. This paper introduces CIGEval, a unified agentic framework for comprehensive evaluation of conditional image generation tasks. CIGEval utilizes large multimodal models (LMMs) as its core, integrating a multi-functional toolbox and establishing a fine-grained evaluation framework. Additionally, we synthesize evaluation trajectories for fine-tuning, empowering smaller LMMs to autonomously select appropriate tools and conduct nuanced analyses based on tool outputs. Experiments across seven prominent conditional image generation tasks demonstrate that CIGEval (GPT-4o version) achieves a high correlation of 0.4625 with human assessments, closely matching the inter-annotator correlation of 0.47. Notably, when implemented with 7B open-source LMMs using only 2.3K training trajectories, CIGEval surpasses the previous GPT-4o-based state-of-the-art method. These findings indicate that CIGEval holds great potential for automating evaluation of image generation tasks while maintaining human-level reliability.

Direct speech translation (ST) has garnered increasing attention nowadays, yet the accurate translation of terminology within utterances remains a great challenge. In this regard, current studies mainly concentrate on leveraging various translation knowledge into ST models. However, these methods often struggle with interference from irrelevant noise and can not fully utilize the translation knowledge. To address these issues, in this paper, we propose a novel Locate-and-Focus method for terminology translation. It first effectively locates the speech clips containing terminologies within the utterance to construct translation knowledge, minimizing irrelevant information for the ST model. Subsequently, it associates the translation knowledge with the utterance and hypothesis from both audio and textual modalities, allowing the ST model to better focus on translation knowledge during translation. Experimental results across various datasets demonstrate that our method effectively locates terminologies within utterances and enhances the success rate of terminology translation, while maintaining robust general translation performance.

Iterative non-autoregressive (NAR) models share a spirit of mixed autoregressive (AR) and fully NAR models, seeking a balance between generation quality and inference efficiency.These models have recently demonstrated impressive performance in varied generation tasks, surpassing the autoregressive Transformer.However, they also face several challenges that impede further development.In this work, we target building more efficient and competitive iterative NAR models.Firstly, we produce two simple metrics to identify the potential problems existing in current refinement processes, and look back on the various iterative NAR models to find the key factors for realizing our purpose.Subsequently, based on the analyses of the limitations of previous inference algorithms, we propose a simple yet effective strategy to conduct efficient refinements without performance declines.Experiments on five widely used datasets show that our final models set the new state-of-the-art performance compared to all previous NAR models, even with fewer decoding steps, and outperform AR Transformer by around one BLEU on average.Our codes and models are available on Github 1 .

BERT-family have been increasingly explored for adaptation to scenarios beyond language understanding tasks, with more recent efforts focused on enabling them to become good instruction followers.These explorations have endowed BERT-family with new roles and human expectations, showcasing their potential on par with current state-of-the-art (SOTA) large language models (LLMs).However, several certain shortcomings in previous BERT-family, such as the relatively sub-optimal training corpora, learning procedure, and model architecture, all impede the further advancement of these models for serving as general and competitive LLMs.Therefore, we aim to address these deficiencies in this paper.Our study not only introduces a more suitable pre-training task that helps BERT-family excel in wider applications to realize generality but also explores the integration of cutting-edge technologies into our model to further enhance their capabilities.Our final models, termed Bidirectional General Language Models (BiGLM), exhibit performance levels comparable to current SOTA LLMs across a spectrum of tasks.Moreover, we conduct detailed analyses to study the effects of scaling and training corpora for BiGLM.To the best of our knowledge, our work represents the early attempt to offer a recipe for building novel types of scalable, general, and competitive LLMs that diverge from current autoregressive modeling methodology.Our codes and models are available on Github 1 .

Bin Xie, Rui Shao, Gongwei Chen, Kaiwen Zhou, Yinchuan Li, Jie Liu, Min Zhang, Liqiang Nie. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Xin Zhang, Ziqi Dai, Yongqi Li, Yanzhao Zhang, Dingkun Long, Pengjun Xie, Meishan Zhang, Jun Yu, Wenjie Li, Min Zhang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Longhui Zhang, Jiahao Wang, Meishan Zhang, GaoXiong Cao, Ensheng Shi, Mayuchi Mayuchi, Jun Yu, Honghai Liu, Jing Li, Min Zhang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Jiaqi Zhao, Miao Zhang, Ming Wang, Yuzhang Shang, Kaihao Zhang, Weili Guan, Yaowei Wang, Min Zhang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Large language models (LLMs) have demonstrated remarkable abilities in various natural language processing areas, but they demand high computation resources which limits their deployment in real-world.Distillation is one technique to solve this problem through either knowledge distillation or task distillation.Both distillation approaches train small models to imitate specific features of LLMs, but they all neglect basic reading education for small models on generic texts that are unrelated to downstream tasks.In this paper, we propose basic reading distillation (BRD) which educates a small model to imitate LLMs basic reading behaviors, such as named entity recognition, question raising and answering, on each sentence.After such basic education, we apply the small model on various tasks including language inference benchmarks and BIG-bench tasks.It shows that the small model can outperform or perform comparable to over 20x bigger LLMs.Analysis reveals that BRD effectively influences the probability distribution of the small model, and has orthogonality to either knowledge distillation or task distillation.

Large Vision-Language Models (LVLMs) have demonstrated remarkable performance across diverse tasks.Despite great success, recent studies show that LVLMs encounter substantial limitations when engaging with visual graphs.To study the reason behind these limitations, we propose VGCURE, a comprehensive benchmark covering 22 tasks for examining the fundamental graph understanding and reasoning capacities of LVLMs.Extensive evaluations conducted on 14 LVLMs reveal that LVLMs are weak in basic graph understanding and reasoning tasks, particularly those concerning relational or structurally complex information.Based on this observation, we propose a structure-aware fine-tuning framework to enhance LVLMs with structure learning abilities through three self-supervised learning tasks.Experiments validate the effectiveness of our method in improving LVLMs' performance on fundamental and downstream graph learning tasks, as well as enhancing their robustness against complex visual graphs.

The efficacy of detectors for texts generated by large language models (LLMs) substantially depends on the availability of large-scale training data. However, white-box zero-shot detectors, which require no such data, are limited by the accessibility of the source model of the LLM-generated text. In this paper, we propose a simple yet effective black-box zero-shot detection approach based on the observation that, from the perspective of LLMs, human-written texts typically contain more grammatical errors than LLM-generated texts. This approach involves calculating the Grammar Error Correction Score (GECScore) for the given text to differentiate between human-written and LLM-generated text. Experimental results show that our method outperforms current state-of-the-art (SOTA) zero-shot and supervised methods, achieving an average AUROC of 98.62% across XSum and Writing Prompts dataset. Additionally, our approach demonstrates strong reliability in the wild, exhibiting robust generalization and resistance to paraphrasing attacks. Data and code are available at: https://github.com/NLP2CT/GECScore.

Generative information extraction (Generative IE) aims to generate structured text sequences from unstructured text using a generative framework. Scaling in model size yields variations in adaptation and generalization, and also drives fundamental shifts in the techniques and approaches used within this domain. In this survey, we first review generative information extraction (IE) methods based on pre-trained language models (PLMs) and large language models (LLMs), focusing on their adaptation and generalization capabilities. We also discuss the connection between these methods and these two aspects. Furthermore, to balance task performance with the substantial computational demands associated with LLMs, we emphasize the importance of model collaboration. Finally, given the advanced capabilities of LLMs, we explore methods for integrating diverse IE tasks into unified models.

Large Language models (LLMs) have become a research hotspot. To accelerate the inference of LLMs, storing computed caches in memory has become the standard technique. However, as the inference length increases, growing KV caches might lead to out-of-memory issues. Many existing methods address this issue through KV cache compression, primarily by preserving key tokens throughout all layers to reduce information loss. Most of them allocate a uniform budget size for each layer to retain. However, we observe that the minimum budget sizes needed to retain essential information vary across layers and models based on the perspectives of attention and hidden state output. Building on this observation, this paper proposes a simple yet effective KV cache compression method that leverages layer uncertainty to allocate budget size for each layer. Experimental results show that the proposed method can reduce memory usage of the KV caches to only ~20% when compared to full KV inference while achieving nearly lossless performance.

Enabling LLMs to handle lengthy context is currently a research hotspot. Most LLMs are built upon rotary position embedding (RoPE), a popular position encoding method. Therefore, a prominent path is to extrapolate the RoPE trained on comparably short texts to far longer texts. A heavy bunch of efforts have been dedicated to boosting the extrapolation via extending the formulations of the RoPE, however, few of them have attempted to showcase their inner workings comprehensively. In this paper, we are driven to offer a straightforward yet in-depth understanding of RoPE extensions from an attention perspective and on two benchmarking tasks. A broad array of experiments reveals several valuable findings: 1) Maintaining attention patterns to those at the pretrained length improves extrapolation; 2) Large attention uncertainty leads to retrieval errors; 3) Using longer continual pretraining lengths for RoPE extensions could reduce attention uncertainty and significantly enhance extrapolation.

Universal Multimodal Retrieval (UMR) aims to enable search across various modalities using a unified model, where queries and candidates can consist of pure text, images, or a combination of both. Previous work has attempted to adopt multimodal large language models (MLLMs) to realize UMR using only text data. However, our preliminary experiments demonstrate that more diverse multimodal training data can further unlock the potential of MLLMs. Despite its effectiveness, the existing multimodal training data is highly imbalanced in terms of modality, which motivates us to develop a training data synthesis pipeline and construct a large-scale, high-quality fused-modal training dataset. Based on the synthetic training data, we develop the General Multimodal Embedder (GME), an MLLM-based dense retriever designed for UMR. Furthermore, we construct a comprehensive UMR Benchmark (UMRB) to evaluate the effectiveness of our approach. Experimental results show that our method achieves state-of-the-art performance among existing UMR methods. Last, we provide in-depth analyses of model scaling and training strategies, and perform ablation studies on both the model and synthetic data.

Large language models (LLMs) have shown remarkable performance in general translation tasks.However, the increasing demand for high-quality translations that are not only adequate but also fluent and elegant.To assess the extent to which current LLMs can meet these demands, we introduce a suitable benchmark (PoetMT) for translating classical Chinese poetry into English.This task requires not only adequacy in translating culturally and historically significant content but also a strict adherence to linguistic fluency and poetic elegance.Our study reveals that existing LLMs fall short of this task.To address these issues, we propose RAT, a Retrieval-Augmented machine Translation method that enhances the translation process by incorporating knowledge related to classical poetry.Additionally, we propose an automatic evaluation metric based on GPT-4, which better assesses translation quality in terms of adequacy, fluency, and elegance, overcoming the limitations of traditional metrics.Our dataset and code will be made available 1 .

High-quality prompts are crucial for eliciting outstanding performance from large language models (LLMs) on complex tasks.Existing research has explored model-driven strategies for prompt optimization.However, these methods often suffer from high computational overhead or require strong optimization capabilities from the model itself, which limits their broad applicability.To address these challenges, we propose ORPP, a framework that enhances model performance by optimizing and generating roleplaying prompts.The core idea of ORPP is to confine the prompt search space to role-playing scenarios, thereby fully activating the model's intrinsic capabilities through carefully crafted, high-quality role-playing prompts.Specifically, ORPP first performs iterative optimization on a small subset of training samples to generate high-quality role-playing prompts.Then, leveraging the model's few-shot learning capability, it transfers the optimization experience to efficiently generate suitable prompts for the remaining samples.Our experimental results show that ORPP not only matches but in most cases surpasses existing mainstream prompt optimization methods in terms of performance.Notably, ORPP suggests great "plug-and-play" capability.In most cases, it can be integrated with various other prompt methods and further enhance their effectiveness.

Hao Fang, Jiawei Kong, Tianqu Zhuang, Yixiang Qiu, Kuofeng Gao, Bin Chen, Shu-Tao Xia, Yaowei Wang, Min Zhang. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Despite the impressive performance of large language models (LLMs) in general domains, they often underperform in specialized domains.Existing approaches typically rely on data synthesis methods and yield promising results by using unlabeled data to capture domain-specific features.However, these methods either incur high computational costs or suffer from performance limitations, while also demonstrating insufficient generalization across different tasks.To address these challenges, we propose AQuilt, a framework for constructing instruction-tuning data for any specialized domains from corresponding unlabeled data, including Answer, Question, Unlabeled data, Inspection, Logic, and Task type.By incorporating logic and inspection, we encourage reasoning processes and self-inspection to enhance model performance.Moreover, customizable task instructions enable high-quality data generation for any task.As a result, we construct a dataset of 703k examples to train a powerful data synthesis model.Experiments show that AQuilt is comparable to DeepSeek-V3 while utilizing just 17% of the production cost.Further analysis demonstrates that our generated data exhibits higher relevance to downstream tasks.

Large language model (LLM) agents typically adopt a step-by-step reasoning framework, in which they interleave the processes of thinking and acting to accomplish the given task.However, this paradigm faces a deeprooted one-pass issue whereby each generated intermediate thought is plugged into the trajectory regardless of its correctness, which can cause irreversible error propagation.To address the issue, this paper proposes a novel framework called Generator-Assistant Stepwise Rollback (GA-Rollback) to induce better decision-making for LLM agents.Particularly, GA-Rollback utilizes a generator to interact with the environment and an assistant to examine each action produced by the generator, where the assistant triggers a rollback operation upon detection of incorrect actions.Moreover, we introduce two additional strategies tailored for the rollback scenario to further improve its effectiveness.Extensive experiments show that GA-Rollback achieves significant improvements over several strong baselines on three widely used benchmarks.Our analysis further reveals that GA-Rollback can function as a robust plug-and-play module, integrating seamlessly with other methods. 1

Recent works have revealed the great potential of speculative decoding in accelerating the autoregressive generation process of large language models.The success of these methods relies on the alignment between draft candidates and the sampled outputs of the target model.Existing methods mainly achieve draft-target alignment with training-based methods, e.g., EAGLE, Medusa, involving considerable training costs.In this paper, we present a trainingfree alignment-augmented speculative decoding algorithm.We propose alignment sampling, which leverages output distribution obtained in the prefilling phase to provide more aligned draft candidates.To further benefit from highquality but non-aligned draft candidates, we also introduce a simple yet effective flexible verification strategy.Through an adaptive probability threshold, our approach can improve generation accuracy while further improving inference efficiency.Experiments on 8 datasets (including question answering, summarization and code completion tasks) show that our approach increases the average generation score by 3.3 points for the LLaMA3 model.Our method achieves a mean acceptance length up to 2.39 and speed up generation by 2.23×.

Enabling digital humans to express rich emotions has significant applications in dialogue systems, gaming, and other interactive scenarios.While recent advances in talking head synthesis have achieved impressive results in lip synchronization, they tend to overlook the rich and dynamic nature of facial expressions.To fill this critical gap, we introduce an endto-end text-to-expression model that explicitly focuses on emotional dynamics.Our model learns expressive facial variations in a continuous latent space and generates expressions that are diverse, fluid, and emotionally coherent.To support this task, we introduce EmoAva, a large-scale and high-quality dataset containing 15,000 text-3D expression pairs.Extensive experiments on both existing datasets and EmoAva demonstrate that our method significantly outperforms baselines across multiple evaluation metrics, marking a significant advancement in the field.

Brain-computer interface (BCI) based on steady- state visual evoked potentials (SSVEP) is a popular paradigm for its simplicity and high information transfer rate (ITR). Accurate and fast SSVEP decoding is crucial for reliable BCI performance. However, conventional decoding methods demand longer time windows, and deep learning models typically require subject-specific fine-tuning, leaving challenges in achieving optimal performance in cross-subject settings. This paper proposed a biofocal masking attention-based method (SSVEP-BiMA) that synergistically leverages the native and symmetric-antisymmetric components for decoding SSVEP. By utilizing multiple signal representations, the network is able to integrate features from a wider range of sample perspectives, leading to more generalized and comprehensive feature learning, which enhances both prediction accuracy and robustness. We performed experiments on two public datasets, and the results demonstrate that our proposed method surpasses baseline approaches in both accuracy and ITR. We believe that this work will contribute to the development of more efficient SSVEP-based BCI systems.

Open Information Extraction (OIE) aims at extracting the relational triplets from open-domain texts. Existing methods, unfortunately, mostly fall prey to the complex OIE setting, due to the failure to extract unseen words and underutilize syntactic features. In this work, we propose a novel system tailored for complex OIE, where a generative PLM with non-autoregressive generative decoding is adopted for abstractive OIE generation. We introduce a heterogeneous syntactic forests as features for the task, merging constituency and dependency forests into a unified syntax graph, aiding in better detecting potential boundaries and relationships within complex terms. Also an Implicit-Explicit Contrastive Learning mechanism is devised to boost the perception of implicit relations and spans. Our system outperforms the current state-of-the-art model across seven OIE datasets.

As large language models (LLMs) are widely applied across various fields, model compression has become increasingly crucial for reducing costs and improving inference efficiency. Post-training pruning is a promising method that does not require resource-intensive iterative training and only needs a small amount of calibration data to assess the importance of parameters. Recent research has enhanced post-training pruning from different aspects but few of them systematically explore the effects of calibration data, and it is unclear if there exist better calibration data construction strategies. We fill this blank and surprisingly observe that calibration data is also crucial to post-training pruning, especially for high sparsity. Through controlled experiments on important influence factors of calibration data, including the pruning settings, the amount of data, and its similarity with pre-training data, we observe that a small size of data is adequate, and more similar data to its pre-training stage can yield better performance. As pre-training data is usually inaccessible for advanced LLMs, we further provide a self-generating calibration data synthesis strategy to construct feasible calibration data. Experimental results on recent strong open-source LLMs (e.g., DCLM, and LLaMA-3) show that the proposed strategy can enhance the performance of strong pruning methods (e.g., Wanda, DSnoT, OWL) by a large margin (up to 2.68%).

Large Language Models~(LLMs) have demonstrated superior performance across a wide range of tasks, but they still exhibit undesirable errors due to incorrect knowledge learned from the training data. To avoid this, knowledge editing methods emerged to precisely edit the specific model knowledge via efficiently modifying a very small percentage of parameters. However, those methods can lead to the problem of **Specificity Failure**, where the existing knowledge and capabilities are severely degraded due to editing. Our preliminary indicates that Specificity Failure primarily stems from the model's attention heads assigning excessive attention scores to entities related to the edited knowledge, thereby unduly focusing on specific snippets within the context, which we denote as the **Attention Drift** phenomenon. To mitigate such Attention Drift issue, we introduce a simple yet effective method **S**elective **A**ttention **D**rift **R**estriction(**SADR**), which introduces an additional regularization term during the knowledge editing process to restrict changes in the attention weight distribution, thereby preventing undue focus on the edited entity. Experiments on five frequently-used strong LLMs demonstrate the effectiveness of our method, where SADR can significantly mitigate Specificity Failure in the predominant knowledge editing tasks.

Large language models (LLMs) have achieved reasonable quality improvements in machine translation (MT). However, most current research on MT-LLMs still faces significant challenges in maintaining translation consistency and accuracy when processing entire documents. In this paper, we introduce DelTA, a Document-levEL Translation Agent designed to overcome these limitations. DelTA features a multi-level memory structure that stores information across various granularities and spans, including Proper Noun Records, Bilingual Summary, Long-Term Memory, and Short-Term Memory, which are continuously retrieved and updated by auxiliary LLM-based components. Experimental results indicate that DelTA significantly outperforms strong baselines in terms of translation consistency and quality across four open/closed-source LLMs and two representative document translation datasets, achieving an increase in consistency scores by up to 4.58 percentage points and in COMET scores by up to 3.16 points on average. DelTA employs a sentence-by-sentence translation strategy, ensuring no sentence omissions and offering a memory-efficient solution compared to the mainstream method. Furthermore, DelTA improves pronoun and context-dependent translation accuracy, and the summary component of the agent also shows promise as a tool for query-based summarization tasks. The code and data of our approach are released at https://github.com/YutongWang1216/DocMTAgent.

Photo-Sharing Multi-modal dialogue generation requires a dialogue agent not only to generate text responses but also to share photos at the proper moment. Using image text caption as the bridge, a pipeline model integrates an image caption model, a text generation model, and an image generation model to handle this complex multi-modal task. However, representing the images with text captions may lose important visual details and information and cause error propagation in the complex dialogue system. Besides, the pipeline model isolates the three models separately because discrete image text captions hinder end-to-end gradient propagation. We propose the first end-to-end model for photo-sharing multi-modal dialogue generation, which integrates an image perceptron and an image generator with a large language model. The large language model employs the vision encoder to perceive visual images in the input end. For image generation in the output end, we propose a dynamic vocabulary transformation matrix and use straight-through and gumbel-softmax techniques to align the large language model and stable diffusion model and achieve end-to-end gradient propagation. We perform experiments on PhotoChat and DialogCC datasets to evaluate our end-to-end model. Compared with pipeline models, the end-to-end model gains state-of-the-art performances on various metrics of text and image generation.

Streaming generation models are utilized across fields, with the Transducer architecture being popular in industrial applications. However, its input-synchronous decoding mechanism presents challenges in tasks requiring non-monotonic alignments, such as simultaneous translation. In this research, we address this issue by integrating Transducer’s decoding with the history of input stream via a learnable monotonic attention. Our approach leverages the forward-backward algorithm to infer the posterior probability of alignments between the predictor states and input timestamps, which is then used to estimate the monotonic context representations, thereby avoiding the need to enumerate the exponentially large alignment space during training. Extensive experiments show that our MonoAttn-Transducer effectively handles non-monotonic alignments in streaming scenarios, offering a robust solution for complex generation tasks. Code is available at https://github.com/ictnlp/MonoAttn-Transducer.

Long-context models (LCMs) have shown great potential in processing long input sequences (even more than 100M tokens) conveniently and effectively. With significant progress, recent research has pointed out that LCMs can accurately locate token-level salient information within the context. Yet, the generation performance of these LCMs is far from satisfactory and might result in misaligned responses, such as hallucinations. To enhance the generation capability of LCMs, existing works have investigated the effects of data size and quality for both pre-training and instruction tuning. Though achieving meaningful improvement, previous methods fall short in either effectiveness or efficiency. In this paper, we introduce LOGO (Long cOntext aliGnment via efficient preference Optimization), a training strategy that first introduces preference optimization for long-context alignment. To overcome the GPU memory-bound issue caused by the long sequence, LOGO employs a reference-free preference optimization strategy and adopts a position synthesis method to construct the training data. By training with only 0.3B data on a single 8 x A800 GPU machine for 16 hours, LOGO allows the Llama-3-8B-Instruct-80K model to achieve comparable performance with GPT-4 in real-world long-context tasks while preserving the model's original capabilities on other tasks, e.g., language modeling and MMLU. Moreover, LOGO can extend the model's context window size while enhancing its generation performance.

Large language models (LLMs) have demonstrated considerable reasoning abilities in various tasks such as mathematics and coding. However, recent studies indicate that even the best models lack true comprehension of their reasoning processes. In this paper, we explore how self-play can enhance the rationality of models in the reasoning process without supervision from humans or superior models. We design a $\textit{\textbf{C}ritic-\textbf{D}iscernment \textbf{G}ame}~(\textbf{CDG})$ in which a prover first provides a solution to a given problem and is subsequently challenged by critiques of its solution. These critiques either aim to assist or mislead the prover. The objective of the prover is to maintain the correct answer when faced with misleading comments, while correcting errors in response to constructive feedback. Our experiments on tasks involving mathematical reasoning, stepwise error detection, self-correction, and long-chain reasoning demonstrate that CDG training can significantly improve the ability of well-aligned LLMs to comprehend their reasoning process.

Large language models (LLMs) have made significant strides in code translation tasks. However, ensuring both the correctness and readability of translated code remains a challenge, limiting their effective adoption in real-world software development. In this work, we propose F2STrans, a function-to-style guiding paradigm designed to progressively improve the performance of LLMs in code translation. Our approach comprises two key stages: (1) Functional learning, which optimizes translation correctness using high-quality source-target code pairs mined from online programming platforms, and (2) Style learning, which improves translation readability by incorporating both positive and negative style examples. Additionally, we introduce a novel code translation benchmark that includes up-to-date source code, extensive test cases, and manually annotated ground-truth translations, enabling comprehensive functional and stylistic evaluations. Experiments on both our new benchmark and existing datasets demonstrate that our approach significantly improves code translation performance. Notably, our approach enables Qwen-1.5B to outperform prompt-enhanced Qwen-32B and GPT-4 on average across 20 diverse code translation scenarios.

Large language models (LLMs) are increasingly being adopted to develop general-purpose AI agents. However, it remains challenging for these LLM-based AI agents to efficiently learn from feedback and iteratively optimize their strategies. To address this challenge, tremendous efforts have been dedicated to designing diverse feedback mechanisms for LLM-based AI agents. To provide a comprehensive overview of this rapidly evolving field, this paper presents a systematic review of these studies, offering a holistic perspective on the feedback mechanisms in LLM-based AI agents. We begin by discussing the construction of LLM-based AI agents, introducing a generalized framework that encapsulates much of the existing work. Next, we delve into the exploration of feedback mechanisms, categorizing them into four distinct types: internal feedback, external feedback, multi-agent feedback, and human feedback. Additionally, we provide an overview of evaluation protocols and benchmarks specifically tailored for LLM-based AI agents. Finally, we highlight the significant challenges and identify potential directions for future studies. The relevant papers are summarized and will be consistently updated at https://github.com/kevinson7515/Agents-Feedback-Mechanisms.

Large language model progress drives code generation research. Most works enhance performance via problem decomposition and knowledge distillation. However, they predominantly adopt a unidirectional approach to knowledge learning. Inspired by the mutually beneficial learning patterns observed in humans, this paper explores a mutual learning paradigm. Specifically, we first employ a dual-training architecture involving two lightweight models that learn the knowledge from each other for better performance. Then, we maximize mutual learning potential by exploring mutual self-correction and preference optimization to address knowledge sharing limits. The self-correction strategy incorporates error information generated by both the model itself and its peer, fostering a mutual refinement process. The preference optimization utilizes the excellent knowledge that has not been learned from the peer model to guide the learning. We conduct and analyze the proposed method on two public datasets, expanding the corpus for optimization. Experimental results show the effectiveness of our method, yielding compelling outcomes that underscore the potential of this collaborative learning approach.

Large Language Models (LLMs) for Generative AI have achieved remarkable progress, evolving into sophisticated and versatile tools widely adopted across various domains and applications. However, the substantial memory overhead caused by their vast number of parameters, combined with the high computational demands of the attention mechanism, poses significant challenges in achieving low latency and high throughput for LLM inference services. Recent advancements, driven by groundbreaking research, have significantly accelerated progress in this field. This paper provides a comprehensive survey of these methods, covering fundamental instance-level approaches, in-depth cluster-level strategies, and emerging scenarios. At the instance level, we review model placement, request scheduling, decoding length prediction, storage management, and the disaggregation paradigm. At the cluster level, we explore GPU cluster deployment, multi-instance load balancing, and cloud service solutions. Additionally, we discuss specific tasks, modules, and auxiliary methods in emerging scenarios. Finally, we outline potential research directions to further advance the field of LLM inference serving.

The 1st Workshop on Cognition-oriented Multimodal Affective and Empathetic Computing (CogMAEC) was held at ACM Multimedia 2025. It focused on moving emotional AI beyond basic recognition toward deeper cognitive understanding. While traditional multimodal affective computing has emphasized simple emotion detection, the rise of multimodal large language models (MLLMs) has spurred interest in modeling how emotions emerge and evolve in context. The workshop gathered researchers on emotion reasoning, multimodal understanding, and human-computer empathy, exploring how machines can not only recognize emotions but also explain their causes and simulate human-like affective reasoning. The program featured invited talks, oral presentations, and posters spanning perception, interaction, causal modeling, and cognitive grounding. CogMAEC provided a platform to connect researchers across disciplines and foster future work on cognitively aware affective computing. Materials are available at https://CogMAEC.github.io/MM2025.

Multi-modal large language models (MLLMs) incorporate heterogeneous modalities into LLMs, enabling a comprehensive understanding of diverse scenarios and objects. Despite the proliferation of evaluation benchmarks and leaderboards for MLLMs, they predominantly overlook the critical capacity of MLLMs to comprehend world knowledge with structured abstractions that appear in visual form. To address this gap, we propose a novel evaluation paradigm and devise M3STR, an innovative benchmark grounded in the Multi-Modal Map for STRuctured understanding. This benchmark leverages multi-modal knowledge graphs to synthesize images encapsulating subgraph architectures enriched with multi-modal entities. M3STR necessitates that MLLMs not only recognize the multi-modal entities within the visual inputs but also decipher intricate relational topologies among them. We delineate the benchmark's statistical profiles and automated construction pipeline, accompanied by an extensive empirical analysis of 26 state-of-the-art MLLMs. Our findings reveal persistent deficiencies in processing abstractive visual information with structured knowledge, thereby charting a pivotal trajectory for advancing MLLMs' holistic reasoning capacities. Code and data are released at https://github.com/zjukg/M3STR

The Ava-MERG Challenge at ACM Multimedia 2025 explores avatar-based multimodal empathetic response generation by advancing dialogue systems that respond with emotional awareness across text, speech, and talking-face video. To support this goal, we constructed a high-quality dataset featuring aligned multimodal responses in diverse conversational scenarios. The challenge comprises two progressively complex subtasks centering around multimodal-aware empathetic response generation. We aim to promote research in empathetic AI and affective computing.

Chen Zhang, Meizhi Zhong, Qimeng Wang, Xuantao Lu, Zheyu Ye, Chengqiang Lu, Yan Gao, Yao Hu, Kehai Chen, Min Zhang, Dawei Song. 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.

Process reward models (PRMs) offer fine-grained, step-level evaluations that facilitate deeper reasoning processes in large language models (LLMs), proving effective in complex tasks like mathematical reasoning. However, developing PRMs is challenging due to the high cost and limited scalability of human-annotated data. Synthetic data from Monte Carlo (MC) estimation is a promising alternative but suffers from a high noise ratio, which can cause overfitting and hinder large-scale training. In this work, we conduct a preliminary study on the noise distribution in synthetic data from MC estimation, identifying that annotation models tend to both underestimate and overestimate step correctness due to limitations in their annotation capabilities. Building on these insights, we propose {\bf S}elf-Denoising Monte {\bf C}arlo {\bf An}notation (\textsc{Scan}), an efficient data synthesis and noise-tolerant learning framework. Our key findings indicate that: (1) Even lightweight models (e.g., 1.5B parameters) can produce high-quality annotations through self-denoising strategy, enabling PRMs to achieve superior performance with only 6\% the inference cost required by vanilla MC estimation. (2) With our robust learning strategy, PRMs can effectively learn from this weak supervision, achieving a 39.2 F1 score improvement (from 19.9 to 59.1) in ProcessBench. Despite using only a compact synthetic dataset, our models surpass strong baselines, including those trained on large-scale human-annotated datasets such as PRM800K. Furthermore, performance continues to improve as we scale up the synthetic data, highlighting the potential of \textsc{Scan} for scalable, cost-efficient, and robust PRM training.

The rapid advancement of Large Language Models (LLMs) has spurred significant progress in Large Speech-Language Models (LSLMs), enhancing their capabilities in both speech understanding and generation. While existing LSLMs often concentrate on augmenting speech generation or tackling a diverse array of short-speech tasks, the efficient processing of long-form speech remains a critical yet underexplored challenge. This gap is primarily attributed to the scarcity of long-speech training datasets and the high computational costs associated with long sequences. To address these limitations, we introduce FastLongSpeech, a novel framework designed to extend LSLM capabilities for efficient long-speech processing without necessitating dedicated long-speech training data. FastLongSpeech incorporates an iterative fusion strategy that can compress excessively long-speech sequences into manageable lengths. To adapt LSLMs for long-speech inputs, it introduces a dynamic compression training approach, which exposes the model to short-speech sequences at varying compression ratios, thereby transferring the capabilities of LSLMs to long-speech tasks. To assess the long-speech capabilities of LSLMs, we develop a long-speech understanding benchmark called LongSpeech-Eval. Experiments show that our method exhibits strong performance in both long-speech and short-speech tasks, while greatly improving inference efficiency.

Large Language Models (LLMs) have demonstrated remarkable instruction-following capabilities across various applications. However, their performance in multilingual settings lacks systematic investigation, with existing evaluations lacking fine-grained constraint analysis across diverse linguistic contexts. We introduce XIFBench, a comprehensive constraint-based benchmark for evaluating multilingual instruction-following abilities of LLMs, comprising 558 instructions with 0-5 additional constraints across five categories (Content, Style, Situation, Format, and Numerical) in six languages spanning different resource levels. To support reliable and consistent cross-lingual evaluation, we implement three methodological innovations: cultural accessibility annotation, constraint-level translation validation, and requirement-based evaluation using English requirements as semantic anchors across languages. Extensive experiments with various LLMs not only quantify performance disparities across resource levels but also provide detailed insights into how language resources, constraint categories, instruction complexity, and cultural specificity influence multilingual instruction-following. Our code and data are available at https://github.com/zhenyuli801/XIFBench.

We introduce \emph{SLED}, an alternative approach to speech language modeling by encoding speech waveforms into sequences of continuous latent representations and modeling them autoregressively using an energy distance objective. The energy distance offers an analytical measure of the distributional gap by contrasting simulated and target samples, enabling efficient training to capture the underlying continuous autoregressive distribution. By bypassing reliance on residual vector quantization, SLED avoids discretization errors and eliminates the need for the complicated hierarchical architectures common in existing speech language models. It simplifies the overall modeling pipeline while preserving the richness of speech information and maintaining inference efficiency. Empirical results demonstrate that SLED achieves strong performance in both zero-shot and streaming speech synthesis, showing its potential for broader applications in general-purpose speech language models. Demos and code are available at \url{https://github.com/ictnlp/SLED-TTS}.

The advancement of Large Language Models (LLMs) has spurred significant interest in Role-Playing Agents (RPAs) for applications such as emotional companionship and virtual interaction. However, recent RPAs are often built on explicit dialogue data, lacking deep, human-like internal thought processes, resulting in superficial knowledge and style expression. While Large Reasoning Models (LRMs) can be employed to simulate character thought, their direct application is hindered by attention diversion (i.e., RPAs forget their role) and style drift (i.e., overly formal and rigid reasoning rather than character-consistent reasoning). To address these challenges, this paper introduces a novel Role-Aware Reasoning (RAR) method, which consists of two important stages: Role Identity Activation (RIA) and Reasoning Style Optimization (RSO). RIA explicitly guides the model with character profiles during reasoning to counteract attention diversion, and then RSO aligns reasoning style with the character and scene via LRM distillation to mitigate style drift. Extensive experiments demonstrate that the proposed RAR significantly enhances the performance of RPAs by effectively addressing attention diversion and style drift.

The rapid development of LLMs has raised concerns about their potential misuse, leading to various watermarking schemes that typically offer high detectability. However, existing watermarking techniques often face trade-off between watermark detectability and generated text quality. In this paper, we introduce Learning to Watermark (LTW), a novel selective watermarking framework that leverages multi-objective optimization to effectively balance these competing goals. LTW features a lightweight network that adaptively decides when to apply the watermark by analyzing sentence embeddings, token entropy, and current watermarking ratio. Training of the network involves two specifically constructed loss functions that guide the model toward Pareto-optimal solutions, thereby harmonizing watermark detectability and text quality. By integrating LTW with two baseline watermarking methods, our experimental evaluations demonstrate that LTW significantly enhances text quality without compromising detectability. Our selective watermarking approach offers a new perspective for designing watermarks for LLMs and a way to preserve high text quality for watermarks. The code is publicly available at: https://github.com/fattyray/learning-to-watermark

This paper introduces **VimoRAG**, a novel video-based retrieval-augmented motion generation framework for motion large language models (LLMs). As motion LLMs face severe out-of-domain/out-of-vocabulary issues due to limited annotated data, **VimoRAG** leverages large-scale in-the-wild video databases to enhance 3D motion generation by retrieving relevant 2D human motion signals. While video-based motion RAG is nontrivial, we address two key bottlenecks: (1) developing an effective motion-centered video retrieval model that distinguishes human poses and actions, and (2) mitigating the issue of error propagation caused by suboptimal retrieval results. We design the Gemini Motion Video Retriever mechanism and the Motion-centric Dual-alignment DPO Trainer, enabling effective retrieval and generation processes. Experimental results show that **VimoRAG** significantly boosts the performance of motion LLMs constrained to text-only input.

Instruction fine-tuning is crucial in NLP tasks, enhancing pretrained models' instruction-following capabilities and task-specific performance. However, obtaining high-quality fine-tuning data for large models is challenging due to data collection difficulties and high production costs. To address this, we propose MASTER, a novel data augmentation method that enriches original data through interactions among multiple agents with varying cognitive levels. We simulate three pedagogically grounded teaching scenarios, leveraging multi-agent conversations to generate high-quality teacher-student interaction data. Utilizing MASTER, we construct BOOST-QA, a fine-tuning dataset augmented from existing datasets like Orca-Math-200k, ProcQA, and OpenHermes2.5. Experiments show that models fine-tuned with BOOST-QA perform excellently across multiple benchmarks, demonstrating strong multitask generalization. Notably, MASTER significantly improves models' reasoning abilities in complex tasks, providing valuable insights for future research.

While large language models (LLMs) demonstrate remarkable success in multilingual translation, their internal core translation mechanisms, even at the fundamental word level, remain insufficiently understood. To address this critical gap, this work introduces a systematic framework for interpreting the mechanism behind LLM translation from the perspective of computational components. This paper first proposes subspace-intervened path patching for precise, fine-grained causal analysis, enabling the detection of components crucial to translation tasks and subsequently characterizing their behavioral patterns in human-interpretable terms. Comprehensive experiments reveal that translation is predominantly driven by a sparse subset of components: specialized attention heads serve critical roles in extracting source language, translation indicators, and positional features, which are then integrated and processed by specific multi-layer perceptrons (MLPs) into intermediary English-centric latent representations before ultimately yielding the final translation. The significance of these findings is underscored by the empirical demonstration that targeted fine-tuning a minimal parameter subset (<5%) enhances translation performance while preserving general capabilities. This result further indicates that these crucial components generalize effectively to sentence-level translation and are instrumental in elucidating more intricate translation tasks.

Searching over semi-structured data with natural language (NL) queries has attracted sustained attention, enabling broader audiences to access information easily. As more applications, such as LLM agents and RAG systems, emerge to search and interact with semi-structured data, two major challenges have become evident: (1) the increasing diversity of domains and schema variations, making domain-customized solutions prohibitively costly; (2) the growing complexity of NL queries, which combine both exact field matching conditions and fuzzy semantic requirements, often involving multiple fields and implicit reasoning. These challenges make formal language querying or keyword-based search insufficient. In this work, we explore neural retrievers as a unified non-formal querying solution by directly index semi-structured collections and understand NL queries. We employ LLM-based automatic evaluation and build a large-scale semi-structured retrieval benchmark (SSRB) using LLM generation and filtering, containing 14M semi-structured objects from 99 different schemas across 6 domains, along with 8,485 test queries that combine both exact and fuzzy matching conditions. Our systematic evaluation of popular retrievers shows that current state-of-the-art models could achieve acceptable performance, yet they still lack precise understanding of matching constraints. While by in-domain training of dense retrievers, the performance can be significantly improved. We believe that our SSRB could serve as a valuable resource for future research in this area, and we hope to inspire further exploration of semi-structured retrieval with complex queries.

Despite rapid advancements in video generation models, generating coherent, long-form storytelling videos that span multiple scenes and characters remains challenging. Current methods often rigidly convert pre-generated keyframes into fixed-length clips, resulting in disjointed narratives and pacing issues. Furthermore, the inherent instability of video generation models means that even a single low-quality clip can significantly degrade the entire output animation’s logical coherence and visual continuity. To overcome these obstacles, we introduce AniMaker, a multi-agent framework enabling efficient multi-candidate clip generation and storytelling-aware clip selection, thus creating globally consistent and story-coherent animation solely from text input. The framework is structured around specialized agents, including the Director Agent for storyboard generation, the Photography Agent for video clip generation, the Reviewer Agent for evaluation, and the Post-Production Agent for editing and voiceover, collectively realizing multi-character, multi-scene animation. Central to AniMaker’s approach are two key technical components: MCTS-Gen in Photography Agent, an efficient Monte Carlo Tree Search (MCTS)-inspired strategy that intelligently navigates the candidate space to generate high-potential clips while optimizing resource usage; and AniEval in Reviewer Agent, the first framework specifically designed for multi-shot animation evaluation, which assesses critical aspects such as story-level consistency, action completion, and animation-specific features by considering each clip in the context of its preceding and succeeding clips. Experiments demonstrate that AniMaker achieves superior quality as measured by popular metrics including VBench and our proposed AniEval framework, while significantly improving the efficiency of multi-candidate generation, pushing AI-generated storytelling animation closer to production standards. Code and data for this paper are at https://animaker-dev.github.io/

Recent advancements in Multimodal Large Language Models (MLLMs) underscore the significance of scalable models and data to boost performance, yet this often incurs substantial computational costs. Although the Mixture of Experts (MoE) architecture has been employed to scale large language or visual-language models efficiently, these efforts typically involve fewer experts and limited modalities. To address this, our work presents the pioneering attempt to develop a unified MLLM with the MoE architecture, named Uni-MoE that can handle a wide array of modalities. Specifically, it features modality-specific encoders with connectors for a unified multimodal representation. We also implement a sparse MoE architecture within the LLMs to enable efficient training and inference through modality-level data parallelism and expert-level model parallelism. To enhance the multi-expert collaboration and generalization, we present a progressive training strategy: 1) Cross-modality alignment using various connectors with different cross-modality data, 2) Training modality-specific experts with cross-modality instruction data to activate experts' preferences, and 3) Tuning the whole Uni-MoE framework utilizing Low-Rank Adaptation (LoRA) on mixed multimodal instruction data. We evaluate the instruction-tuned Uni-MoE on a comprehensive set of multimodal datasets. The extensive experimental results demonstrate Uni-MoE's principal advantage of significantly reducing performance bias in handling mixed multimodal datasets, alongside improved multi-expert collaboration and generalization.

As a cross-topic of multi-view learning and multi-label classification, multi-view multi-label classification has gradually gained traction in recent years. The application of multi-view contrastive learning has further facilitated this process; however, the existing multi-view contrastive learning methods crudely separate the so-called negative pair, which largely results in the separation of samples belonging to the same category or similar ones. Besides, plenty of multi-view multi-label learning methods ignore the possible absence of views and labels. To address these issues, in this paper, we propose an incomplete multi-view missing multi-label classification network named RANK. In this network, a label-driven multi-view contrastive learning strategy is proposed to leverage supervised information to preserve the intra-view structure and perform the cross-view consistency alignment. Furthermore, we break through the view-level weights inherent in existing methods and propose a quality-aware subnetwork to dynamically assign quality scores to each view of each sample. The label correlation information is fully utilized in the final multi-label cross-entropy classification loss, effectively improving the discriminative power. Last but not least, our model is not only able to handle complete multi-view multi-label data, but also works on datasets with missing instances and labels. Extensive experiments confirm that our RANK outperforms existing state-of-the-art methods.

Abstract Autoregressive language models demonstrate excellent performance in various scenarios. However, the inference efficiency is limited by its one-step-one-word generation mode, which has become a pressing problem recently as the models become increasingly larger. Speculative decoding employs a “draft and then verify” mechanism to allow multiple tokens to be generated in one step, realizing lossless acceleration. Existing methods mainly adopt fixed heuristic draft structures, which do not adapt to different situations to maximize the acceptance length during verification. To alleviate this dilemma, we propose OPT-Tree, an algorithm to construct adaptive and scalable draft trees, which can be applied to any autoregressive draft model. It searches the optimal tree structure that maximizes the mathematical expectation of the acceptance length in each decoding step. Experimental results reveal that OPT-Tree outperforms the existing draft structures and achieves a speed-up ratio of up to 3.2 compared with autoregressive decoding. If the draft model is powerful enough and the node budget is sufficient, it can generate more than ten tokens in a single step. Our code is available at https://github.com/Jikai0Wang/OPT-Tree.

Cross-domain Aspect Sentiment Triplet Extraction (ASTE) aims to extract the triplets from the review of a target domain, utilizing knowledge from a source domain. As a newly proposed task, limited work has been devoted to it. Except for solving it in a zero-shot manner with in-domain models, recent work explores a bidirectional generative framework to generate pseudo-labeled target data. However, such a method suffers from low efficiency with two-stage training and unstable pseudo-label quality. In this paper, we propose a Hybrid Prompts Mixture (HiPM) method for cross-domain ASTE to fully utilize domain-independent knowledge. Within this method, given that syntax information is an essential linguistic feature for triplet extraction, we design a syntax-related hard prompt to transfer the structures. Additionally, aspects from different domains exhibit similarities in their respective categories. We take this shared information as the prototypes and enrich them through a warm-up step. The resulting prototypes then act as the source of soft prompts. We further mix the hard and soft prompts with the original sequence into a generative model to extract triplets. Experimental results show that our method outperforms baselines on twelve transfer pairs, and obtains a 1.48% average F1 score improvement over the state-of-the-art cross-domain ASTE model.

Electroencephalography (EEG) is an objective tool for emotion recognition with promising applications. However, the scarcity of labeled data remains a major challenge in this field, limiting the widespread use of EEG-based emotion recognition. In this paper, a semi-supervisedDual-streamSelf-attentiveAdversarialGraphContrastive learning framework (termed asDS-AGC) is proposed to tackle the challenge of limited labeled data in cross-subject EEG-based emotion recognition. The DS-AGC framework includes two parallel streams for extracting non-structural and structural EEG features. The non-structural stream incorporates a semi-supervised multi-domain adaptation method to alleviate distribution discrepancy among labeled source domain, unlabeled source domain, and unknown target domain. The structural stream develops a graph contrastive learning method to extract effective graph-based feature representation from multiple EEG channels in a semi-supervised manner. Further, a self-attentive fusion module is developed for feature fusion, sample selection, and emotion recognition, which highlights EEG features more relevant to emotions and data samples in the labeled source domain that are closer to the target domain. Extensive experiments are conducted on four benchmark databases (SEED, SEED-IV, SEED-V, and FACED) using a semi-supervised cross-subject leave-one-subject-out cross-validation evaluation protocol. The results show that the proposed model outperforms existing methods under different incomplete label conditions with an average improvement of 2.17%, which demonstrates its effectiveness in addressing the label scarcity problem in cross-subject EEG-based emotion recognition.

Dependency parsing is a fundamental task in natural language processing that involves identifying the grammatical relationships between words in a sentence. One promising approach for performing this task in languages lacking annotated treebanks is treebank translation, which utilizes word alignments to map dependencies from a source treebank to the corresponding target translation. However, due to language differences and the limitations of word alignment tools, this method would inevitably generate noise during mapping. To reduce the effect of noise, we first exploit MetaNet to compute quality scores for each dependency and identify low-score ones as noise. MetaNet is a fake teacher that learns to score homework (dependencies) by comparing answers from the top student (strong parser) and the regular student (weak parser) without knowing the correct answer (gold-standard). With the scoring capability of MetaNet, we design an iterative algorithm to boost the target treebank quality, which trains with high-quality dependencies and relabels the low-quality dependencies. Our method achieves better results than the originally translated treebanks and shows highly competitive performances with prior methods on the Universal Dependency Treebanks v2.2. We also provide detailed analysis and discussions.

Text-based person search (TBPS) aims to retrieve images of a specific person from a large image gallery based on a natural language description. Existing methods rely on massive annotated image-text data to achieve satisfactory performance in fully-supervised learning. This presents a substantial practical challenge, given the difficulty in obtaining annotated texts for person images. This work undertakes a pioneering initiative to explore TBPS under the semi-supervised setting, where only a limited number of person images are annotated with textual descriptions while the majority of images lack annotations. We present a two-stage basic solution based on generation-then-retrieval for semi-supervised TBPS. The generation stage enriches annotated data by applying an image captioning model to generate pseudo-texts for unannotated images. Later, the retrieval stage performs fully-supervised retrieval learning using the augmented data. Crucially, considering the noise interference of the pseudo-texts on retrieval learning, we propose a noise-robust retrieval framework that enhances the ability of the retrieval model to handle noisy data. The framework integrates two key strategies: Hybrid Patch-Channel Masking (PC-Mask) to refine the model architecture, and Noise-Guided Progressive Training (NP-Train) to enhance the training process. PC-Mask performs masking on the input data at both the patch-level and the channel-level to prevent overfitting noisy supervision. NP-Train introduces a progressive training schedule based on the noise level of pseudo-texts to facilitate noise-robust learning. Extensive experiments on multiple TBPS benchmarks show that the proposed framework achieves promising performance under the semi-supervised setting.

Distantly supervised named entity recognition (DS-NER) has emerged as a cheap and convenient alternative to traditional human annotation methods, enabling the automatic generation of training data by aligning text with external resources. Despite the many efforts in noise measurement methods, few works focus on the latent noise distribution between different distant annotation methods. In this work, we explore the effectiveness and robustness of DS-NER by two aspects: (1) distant annotation techniques, which encompasses both traditional rule-based methods and the innovative large language model supervision approach, and (2) noise assessment, for which we introduce a novel framework. This framework addresses the challenges by distinctly categorizing them into theunlabeled-entity problem (UEP)and thenoisy-entity problem (NEP), subsequently providing specialized solutions for each. Our proposed method achieves significant improvements on eight real-world distant supervision datasets originating from three different data sources and involving four distinct annotation techniques, confirming its superiority over current state-of-the-art methods.

Text-based Person Search (TBPS) aims to retrieve the person images using natural language descriptions. Recently, Contrastive Language Image Pretraining (CLIP), a universal large cross-modal vision-language pre-training model, has remarkably performed over various cross-modal downstream tasks due to its powerful cross-modal semantic learning capacity. TPBS, as a fine-grained cross-modal retrieval task, is also facing the rise of research on the CLIP-based TBPS. In order to explore the potential of the visual-language pre-training model for downstream TBPS tasks, this paper makes the first attempt to conduct a comprehensive empirical study of CLIP for TBPS and thus contribute a straightforward, incremental, yet strong TBPS-CLIP baseline to the TBPS community. We revisit critical design considerations under CLIP, including data augmentation and loss function. The model, with the aforementioned designs and practical training tricks, can attain satisfactory performance without any sophisticated modules. Also, we conduct the probing experiments of TBPS-CLIP in model generalization and model compression, demonstrating the effectiveness of TBPS-CLIP from various aspects. This work is expected to provide empirical insights and highlight future CLIP-based TBPS research.

Large language models (LLMs) have been widely used in various applications but are known to suffer from issues related to untruthfulness and toxicity. While parameter-efficient modules (PEMs) have demonstrated their effectiveness in equipping models with new skills, leveraging PEMs for deficiency unlearning remains underexplored. In this work, we propose a PEMs operation approach, namely Extraction-before-Subtraction (Ext-Sub), to enhance the truthfulness and detoxification of LLMs through the integration of ``expert'' PEM and ``anti-expert'' PEM. Remarkably, even anti-expert PEM possess valuable capabilities due to their proficiency in generating fabricated content, which necessitates language modeling and logical narrative competence. Rather than merely negating the parameters, our approach involves extracting and eliminating solely the deficiency capability within anti-expert PEM while preserving the general capabilities. To evaluate the effectiveness of our approach in terms of truthfulness and detoxification, we conduct extensive experiments on LLMs, encompassing additional abilities such as language modelling and mathematical reasoning. Our empirical results demonstrate that our approach effectively improves truthfulness and detoxification, while largely preserving the fundamental abilities of LLMs.

In this study, we address the challenge of enhancing temporal knowledge reasoning in Large Language Models (LLMs).LLMs often struggle with this task, leading to the generation of inaccurate or misleading responses.This issue mainly arises from their limited ability to handle evolving factual knowledge and complex temporal logic.To overcome these limitations, we propose Abstract Reasoning Induction (ARI) framework, which divides temporal reasoning into two distinct phases: Knowledgeagnostic and Knowledge-based.This framework offers factual knowledge support to LLMs while minimizing the incorporation of extraneous noisy data.Concurrently, informed by the principles of constructivism, ARI provides LLMs the capability to engage in proactive, self-directed learning from both correct and incorrect historical reasoning samples.By teaching LLMs to actively construct knowledge and methods, it can significantly boosting their temporal reasoning abilities.Our approach achieves significant improvements, with relative gains of 29.7% and 9.27% on two temporal QA datasets, underscoring its efficacy in advancing temporal reasoning in LLMs.The code can be found at https: //github.com/czy1999/ARI-QA.

Two-pass direct speech-to-speech translation (S2ST) models have shown promising results which decompose S2ST into speech-to-text translation (S2TT) and text-to-speech (TTS), yet conduct end-to-end training by sharing the target text representation between S2TT and TTS models.However, the training of these models still requires large-scale parallel speech data comprising triplets, which is extremely challenging to collect.On the other hand, S2TT and TTS have accumulated a large amount of data and numerous pretrained models, which can be used to reduce the reliance on parallel speech data.To this end, we propose a composite S2ST model named ComSpeech, which connects pretrained S2TT and TTS models by introducing a vocabulary adaptor based on connectionist temporal classification (CTC).The vocabulary adaptor is employed to adapt the output text sequence of S2TT to the input text sequence of TTS, which are different due to the use of different vocabularies.In this way, ComSpeech can still be trained end-to-end and only needs a small amount of parallel speech data to finetune.We further propose a novel training method ComSpeech-ZS to eliminate the reliance on parallel speech data by aligning the text representation space of S2TT and TTS.Experimental results on the CVSS dataset show that when the parallel speech data is available, ComSpeech surpasses previous two-pass models like UnitY and Translatotron 2 in both translation quality and decoding speed.When there is no parallel speech data, ComSpeech-ZS lags behind ComSpeech by only 0.7 ASR-BLEU and outperforms the cascaded models. 1

Multimodal reasoning with large language models (LLMs) often suffers from hallucinations and the presence of deficient or outdated knowledge within LLMs.Some approaches have sought to mitigate these issues by employing textual knowledge graphs, but their singular modality of knowledge limits comprehensive cross-modal understanding.In this paper, we propose the Multimodal Reasoning with Multimodal Knowledge Graph (MR-MKG) method, which leverages multimodal knowledge graphs (MMKGs) to learn rich and semantic knowledge across modalities, significantly enhancing the multimodal reasoning capabilities of LLMs.In particular, a relation graph attention network is utilized for encoding MMKGs and a cross-modal alignment module is designed for optimizing image-text alignment.A MMKGgrounded dataset is constructed to equip LLMs with initial expertise in multimodal reasoning through pretraining.Remarkably, MR-MKG achieves superior performance while training on only a small fraction of parameters, approximately 2.25% of the LLM's parameter size.Experimental results on multimodal question answering and multimodal analogy reasoning tasks demonstrate that our MR-MKG method outperforms previous state-of-the-art models.

Evaluating and Rethinking the current landscape of Large Multimodal Models (LMMs), we observe that widely-used visual-language projection approaches (e.g., Q-former or MLP) focus on the alignment of image-text descriptions yet ignore the visual knowledgedimension alignment, i.e., connecting visuals to their relevant knowledge.Visual knowledge plays a significant role in analyzing, inferring, and interpreting information from visuals, helping improve the accuracy of answers to knowledge-based visual questions.In this paper, we mainly explore improving LMMs with visual-language knowledge alignment, especially aimed at challenging knowledge-based visual question answering (VQA).To this end, we present a Cognitive Visual-Language Mapper (CVLM), which contains a pretrained Visual Knowledge Aligner (VKA) and a Finegrained Knowledge Adapter (FKA) used in the multimodal instruction tuning stage.Specifically, we design the VKA based on the interaction between a small language model and a visual encoder, training it on collected imageknowledge pairs to achieve visual knowledge acquisition and projection.FKA is employed to distill the fine-grained visual knowledge of an image and inject it into Large Language Models (LLMs).We conduct extensive experiments on knowledge-based VQA benchmarks and experimental results show that CVLM significantly improves the performance of LMMs on knowledge-based VQA (average gain by 5.0%).Ablation studies also verify the effectiveness of VKA and FKA, respectively.1

Simultaneous translation models play a crucial role in facilitating communication.However, existing research primarily focuses on text-totext or speech-to-text models, necessitating additional cascade components to achieve speechto-speech translation.These pipeline methods suffer from error propagation and accumulate delays in each cascade component, resulting in reduced synchronization between the speaker and listener.To overcome these challenges, we propose a novel non-autoregressive generation framework for simultaneous speech translation (NAST-S2x 1 ), which integrates speechto-text and speech-to-speech tasks into a unified end-to-end framework.We develop a nonautoregressive decoder capable of concurrently generating multiple text or acoustic unit tokens upon receiving fixed-length speech chunks.The decoder can generate blank or repeated tokens and employ CTC decoding to dynamically adjust its latency.Experimental results show that NAST-S2x outperforms state-of-theart models in both speech-to-text and speechto-speech tasks.It achieves high-quality simultaneous interpretation within a delay of less than 3 seconds and provides a 28× decoding speedup in offline generation. 2

Keqin Peng, Liang Ding, Yancheng Yuan, Xuebo Liu, Min Zhang, Yuanxin Ouyang, Dacheng Tao. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Zhaochen Su, Juntao Li, Jun Zhang, Tong Zhu, Xiaoye Qu, Pan Zhou, Yan Bowen, Yu Cheng, Min Zhang. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Reconstructing natural language from noninvasive electroencephalography (EEG) holds great promise as a language decoding technology for brain-computer interfaces (BCIs).How-

Large language models (LLMs) have exhibited remarkable performance in various natural language processing tasks.Techniques like instruction tuning have effectively enhanced the proficiency of LLMs in the downstream task of machine translation.However, the existing approaches fail to yield satisfactory translation outputs that match the quality of supervised neural machine translation (NMT) systems.One plausible explanation for this discrepancy is that the straightforward prompts employed in these methodologies are unable to fully exploit the acquired instruction-following capabilities.To this end, we propose the TASTE framework, which stands for translating through selfreflection.The self-reflection process includes two stages of inference.In the first stage, LLMs are instructed to generate preliminary translations and conduct self-assessments on these translations simultaneously.In the second stage, LLMs are tasked to refine these preliminary translations according to the evaluation results.The evaluation results in four language directions on the WMT22 benchmark reveal the effectiveness of our approach compared to existing methods.Our work presents a promising approach to unleash the potential of LLMs and enhance their capabilities in MT.The codes and datasets are open-sourced at https://github. com/YutongWang1216/ReflectionLLMMT.

End-to-end speech translation (ST) presents notable disambiguation challenges as it necessitates simultaneous cross-modal and crosslingual transformations.While word sense disambiguation is an extensively investigated topic in textual machine translation, the exploration of disambiguation strategies for ST models remains limited.Addressing this gap, this paper introduces the concept of speech sense disambiguation (SSD), specifically emphasizing homophones -words pronounced identically but with different meanings.To facilitate this, we first create a comprehensive homophone dictionary and an annotated dataset rich with homophone information established based on speech-text alignment.Building on this unique dictionary, we introduce AmbigST, an innovative homophone-aware contrastive learning approach that integrates a homophone-aware masking strategy.Our experiments on different MuST-C and CoVoST ST benchmarks demonstrate that AmbigST sets new performance standards.Specifically, it achieves SOTA results on BLEU scores for English to German, Spanish, and French ST tasks, underlining its effectiveness in reducing speech sense ambiguity.

Simultaneous speech-to-speech translation (Simul-S2ST, a.k.a streaming speech translation) outputs target speech while receiving streaming speech inputs, which is critical for real-time communication.Beyond accomplishing translation between speech, Simul-S2ST requires a policy to control the model to generate corresponding target speech at the opportune moment within speech inputs, thereby posing a double challenge of translation and policy.In this paper, we propose StreamSpeech, a direct Simul-S2ST model that jointly learns translation and simultaneous policy in a unified framework of multi-task learning.Adhering to a multi-task learning approach, StreamSpeech can perform offline and simultaneous speech recognition, speech translation and speech synthesis via an "All-in-One" seamless model.Experiments on CVSS benchmark demonstrate that StreamSpeech achieves state-of-the-art performance in both offline S2ST and Simul-S2ST tasks.Besides, StreamSpeech is able to present high-quality intermediate results (i.e., ASR or translation results) during simultaneous translation process, offering a more comprehensive real-time communication experience 1 .

Simultaneous Machine Translation (SiMT) aims to yield a real-time partial translation with a monotonically growing source-side context.However, there is a counterintuitive phenomenon about the context usage between training and inference: e.g., in wait-k inference, model consistently trained with wait-k is much worse than that model inconsistently trained with wait-k ′ (k ′ ̸ = k) in terms of translation quality.To this end, we first investigate the underlying reasons behind this phenomenon and uncover the following two factors: 1) the limited correlation between translation quality and training loss; 2) exposure bias between training and inference.Based on both reasons, we then propose an effective training approach called context consistency training accordingly, which encourages consistent context usage between training and inference by optimizing translation quality and latency as bi-objectives and exposing the predictions to the model during the training.The experiments on three language pairs demonstrate that our SiMT system encouraging context consistency outperforms existing SiMT systems with context inconsistency for the first time.1

End-to-end automatic speech recognition (ASR) systems have made significant progress in general scenarios.However, it remains challenging to transcribe contextual named entities (NEs) in the contextual ASR scenario.Previous approaches have attempted to address this by utilizing the NE dictionary.These approaches treat entities as individual tokens and generate them token-by-token, which may result in incomplete transcriptions of entities.In this paper, we treat entities as indivisible wholes and introduce the idea of copying into ASR.We design a systematic mechanism called CopyNE, which can copy entities from the NE dictionary.By copying all tokens of an entity at once, we can reduce errors during entity transcription, ensuring the completeness of the entity.Experiments demonstrate that CopyNE consistently improves the accuracy of transcribing entities compared to previous approaches.Even when based on the strong Whisper, CopyNE still achieves notable improvements.

Nowadays, character-based sequence labeling becomes the mainstream Chinese named entity recognition (CNER) approach, instead of word-based methods, since the latter degrades performance due to propagation of word segmentation (WS) errors. To make use of WS information, previous studies usually learn CNER and WS simultaneously with multi-task learning (MTL) framework, or treat WS information as extra guide features for CNER model, in which the utilization of WS information is indirect and shallow. In light of the complementary information inside multi-grained words, and the close connection between named entities and part-of-speech (POS) tags, this work proposes a tree parsing approach for joint modeling CNER, multi-grained word segmentation (MWS) and POS tagging tasks simultaneously. Specifically, we first propose a unified tree representation for MWS, POS tagging, and CNER.Then, we automatically construct the MWS-POS-NER data based on the unified tree representation for model training. Finally, we present a two-stage joint tree parsing framework. Experimental results on OntoNotes4 and OntoNotes5 show that our proposed approach of jointly modeling CNER with MWS and POS tagging achieves better or comparable performance with latest methods.

Large Language Models (LLMs) have demonstrated impressive performances across various NLP tasks with just a few prompts via in-context learning. Previous studies have emphasized the pivotal role of well-chosen examples in in-context learning, as opposed to randomly selected instances that exhibits unstable results.A successful example selection scheme depends on multiple factors, while in the context of LLMs-based machine translation, the common selection algorithms only consider the single factor, i.e., the similarity between the example source sentence and the input sentence.In this paper, we introduce a novel approach to use multiple translational factors for in-context example selection by using monotone submodular function maximization.The factors include surface/semantic similarity between examples and inputs on both source and target sides, as well as the diversity within examples.Importantly, our framework mathematically guarantees the coordination between these factors, which are different and challenging to reconcile.Additionally, our research uncovers a previously unexamined dimension: unlike other NLP tasks, the translation part of an example is also crucial, a facet disregarded in prior studies.Experiments conducted on BLOOMZ-7.1B and LLAMA2-13B, demonstrate that our approach significantly outperforms random selection and robust single-factor baselines across various machine translation tasks.

In this paper, we propose a new setting for generating product descriptions from images, augmented by marketing keywords. It leverages the combined power of visual and textual information to create descriptions that are more tailored to the unique features of products. For this setting, previous methods utilize visual and textual encoders to encode the image and keywords and employ a language model-based decoder to generate the product description. However, the generated description is often inaccurate and generic since same-category products have similar copy-writings, and optimizing the overall framework on large-scale samples makes models concentrate on common words yet ignore the product features. To alleviate the issue, we present a simple and effective Multimodal In-Context Tuning approach, named ModICT, which introduces a similar product sample as the reference and utilizes the in-context learning capability of language models to produce the description. During training, we keep the visual encoder and language model frozen, focusing on optimizing the modules responsible for creating multimodal in-context references and dynamic prompts. This approach preserves the language generation prowess of large language models (LLMs), facilitating a substantial increase in description diversity. To assess the effectiveness of ModICT across various language model scales and types, we collect data from three distinct product categories within the E-commerce domain. Extensive experiments demonstrate that ModICT significantly improves the accuracy (by up to 3.3% on Rouge-L) and diversity (by up to 9.4% on D-5) of generated results compared to conventional methods. Our findings underscore the potential of ModICT as a valuable tool for enhancing the automatic generation of product descriptions in a wide range of applications. Data and code are at https://github.com/HITsz-TMG/Multimodal-In-Context-Tuning

Multimodal machine translation (MMT) is a challenging task that seeks to improve translation quality by incorporating visual information. However, recent studies have indicated that the visual information provided by existing MMT datasets is insufficient, causing models to disregard it and overestimate their capabilities. This issue presents a significant obstacle to the development of MMT research. This paper presents a novel solution to this issue by introducing 3AM, an ambiguity-aware MMT dataset comprising 26,000 parallel sentence pairs in English and Chinese, each with corresponding images. Our dataset is specifically designed to include more ambiguity and a greater variety of both captions and images than other MMT datasets. We utilize a word sense disambiguation model to select ambiguous data from vision-and-language datasets, resulting in a more challenging dataset. We further benchmark several state-of-the-art MMT models on our proposed dataset. Experimental results show that MMT models trained on our dataset exhibit a greater ability to exploit visual information than those trained on other MMT datasets. Our work provides a valuable resource for researchers in the field of multimodal learning and encourages further exploration in this area. The data, code and scripts are freely available at https://github.com/MaxyLee/3AM.

Multimodal Entity Linking (MEL) is the task of mapping mentions with multimodal contexts to the referent entities from a knowledge base. Existing MEL methods mainly focus on designing complex multimodal interaction mechanisms and require fine-tuning all model parameters, which can be prohibitively costly and difficult to scale in the era of Large Language Models (LLMs). In this work, we propose GEMEL, a Generative Multimodal Entity Linking framework based on LLMs, which directly generates target entity names. We keep the vision and language model frozen and only train a feature mapper to enable cross-modality interactions. To adapt LLMs to the MEL task, we leverage the in-context learning capability of LLMs by retrieving multimodal instances as demonstrations. Extensive experiments show that, with only ∼0.3% of the model parameters fine-tuned, GEMEL achieves state-of-the-art results on two well-established MEL datasets (7.7% accuracy gains on WikiDiverse and 8.8% accuracy gains on WikiMEL). The performance gain stems from mitigating the popularity bias of LLM predictions and disambiguating less common entities effectively. Further analysis verifies the generality and scalability of GEMEL. Our framework is compatible with any off-the-shelf language model, paving the way towards an efficient and general solution for utilizing LLMs in the MEL task. Our code is available at https://github.com/HITsz-TMG/GEMEL.

Recent generative large language models (LLMs) have exhibited incredible instruction-following capabilities while keeping strong task completion ability, even without task-specific fine-tuning. Some works attribute this to the bonus of the new scaling law, in which the continuous improvement of model capacity yields emergent capabilities, e.g., reasoning and universal generalization. However, we point out that recent LLMs still show shortcut learning behavior, where the models tend to exploit spurious correlations between non-robust features and labels for prediction, which might lead to overestimating model capabilities. LLMs memorize more complex spurious correlations (i.e., task \leftrightarrow feature \leftrightarrow label) compared with that learned from previous pre-training and task-specific fine-tuning paradigm (i.e., feature \leftrightarrow label). Based on our findings, we propose FSLI, a framework for encouraging LLMs to Forget Spurious correlations and Learn from In-context information. Experiments on three tasks show that FSFI can effectively mitigate shortcut learning. Besides, we argue not to overestimate the capabilities of LLMs and conduct evaluations in more challenging and complete test scenarios.

Large Language Models (LLMs) hold considerable promise for artificial general intelligence, given their intrinsic abilities to accomplish a wide range of open-domain tasks either independently or in tandem with specialized expert models. However, despite these capabilities, the performance of LLMs has yet to be comprehensively evaluated in realistic scenarios. To this end, in this work, we introduce a novel task, the Realistic Chinese Spell Checking (RCSC), to evaluate the effectiveness of existing methods comprehensively. In contrast to existing works that solely address Chinese character misspellings or pinyin conversions, our task aims to convert the realistic Chinese text into the corresponding correct text. The realistic Chinese text may potentially contain both Chinese misspellings and pinyin conversions. We first present the Realistic Chinese Spell Checking Benchmark (RCSCB), which consists of two subsets and contains a total of 581,657 samples. Then, we benchmark the performance of various baselines and find that all the existing methods, including instruction-based LLMs, achieve unsatisfactory results on RCSCB. To further improve the performance on RCSCB, we propose Pinyin-Enhanced Spell Checker (PESC), which is specifically designed to address pinyin-related misspellings. Experimental results demonstrate that PESC can achieve state-of-the-art performance on RCSCB. Despite the progress made, the current state-of-the-art performance is still far from satisfactory. We expect further progress on this crucial and challenging task.

Chinese sequence labeling tasks are sensitive to word boundaries. Although pretrained language models (PLM) have achieved considerable success in these tasks, current PLMs rarely consider boundary information explicitly. An exception to this is BABERT, which incorporates unsupervised statistical boundary information into Chinese BERT’s pre-training objectives. Building upon this approach, we input supervised high-quality boundary information to enhance BABERT’s learning, developing a semi-supervised boundary-aware PLM. To assess PLMs’ ability to encode boundaries, we introduce a novel “Boundary Information Metric” that is both simple and effective. This metric allows comparison of different PLMs without task-specific fine-tuning. Experimental results on Chinese sequence labeling datasets demonstrate that the improved BABERT version outperforms the vanilla version, not only in these tasks but also in broader Chinese natural language understanding tasks. Additionally, our proposed metric offers a convenient and accurate means of evaluating PLMs’ boundary awareness.

In the field of Natural Language Processing (NLP), Aspect-Based Sentiment Analysis (ABSA) has gained significant attention in recent years due to its ability to perform fine-grained sentiment analysis. Generative methods tackle various ABSA tasks by autoregressively generating the target sequence of sentiment tuples in a specified format. However, the sentiment tuple is intrinsically an unordered set, and the method introduces an order bias between the generated sequence and the original target. Therefore, to investigate the impact of sentiment tuples order on model performance, we conduct a pilot experiment, unveiling that the order of tuples significantly influences the learning outcomes of the Seq2Seq model. Thus, we propose a novel tuple-order learning method that prioritizes tuples from simple to complex, facilitated by a discrete evaluation method that assesses the difficulty of each individual tuple. Specifically, we incorporate positional information on tuples and employ an effective strategy to expedite the assessment of individual tuples. The method optimizes the learning process while maintaining the structural integrity of existing generative models. Extensive experiments show that our approach significantly advances the performance on 14 datasets of 5 benchmark tasks. We will release our code at https://github.com/gongzhenhu/TOL.

This work proposes a simple training-free prompt-free approach to leverage large language models (LLMs) for the Chinese spelling correction (CSC) task, which is totally different from all previous CSC approaches.The key idea is to use an LLM as a pure language model in a conventional manner.The LLM goes through the input sentence from the beginning, and at each inference step, produces a distribution over its vocabulary for deciding the next token, given a partial sentence.To ensure that the output sentence remains faithful to the input sentence, we design a minimal distortion model that utilizes pronunciation or shape similarities between the original and replaced characters.Furthermore, we propose two useful reward strategies to address practical challenges specific to the CSC task.Experiments on five public datasets demonstrate that our approach significantly improves LLM performance, enabling them to compete with state-of-the-art domain-general CSC models.

Large language models (LLMs) often exhibit excessive, random, and uninformative uncertainty, rendering them unsuitable for decisionmaking in human-computer interactions.In this paper, we aim to instigate a heightened awareness of self-uncertainty in LLMs, enabling them to express uncertainty more effectively.To accomplish this, we propose an uncertainty-aware instruction tuning (UaIT) method, aligning LLMs' perception with the probabilistic uncertainty of the generation.We conducted experiments using LLaMA2 and Mistral on multiple free-form QA tasks.Experimental results revealed a surprising 45.2% improvement in the effectiveness of uncertainty expression by LLMs, accompanied by reasonably good out-of-domain generalization capabilities.Moreover, this uncertainty expression can serve as a valuable real-time basis for human decision-making, e.g., retrieving external documents and incorporating stronger LLMs 1 .

Consistency learning (CL) has proven to be a valuable technique for improving the robustness of models in conditional sentence generation (CSG) tasks by ensuring stable predictions across various input data forms.However, models augmented with CL often face challenges in optimizing consistency features, which can detract from their efficiency and effectiveness.To address these challenges, we introduce Curriculum Consistency Learning (CCL), a novel strategy that guides models to learn consistency in alignment with their current capacity to differentiate between features.CCL is designed around the inherent aspects of CL-related losses, promoting task independence and simplifying implementation.Implemented across four representative CSG tasks, including instruction tuning (IT) for large language models and machine translation (MT) in three modalities (text, speech, and vision), CCL demonstrates marked improvements.Specifically, it delivers +2.0 average accuracy point improvement compared with vanilla IT and an average increase of +0.7 in COMET scores over traditional CL methods in MT tasks.Our comprehensive analysis further indicates that models utilizing CCL are particularly adept at managing complex instances, showcasing the effectiveness and efficiency of CCL in improving CSG models.Code and scripts are available at https://github.com/xinxinxing/ Curriculum-Consistency-Learning.

Recent studies have explored the working mechanisms of In-Context Learning (ICL).However, they mainly focus on classification and simple generation tasks, limiting their broader application to more complex generation tasks in practice.To address this gap, we investigate the impact of demonstrations on token representations within the practical alignment tasks.We find that the transformer embeds the task function learned from demonstrations into the separator token representation, which plays an important role in the generation of prior response tokens.Once the prior response tokens are determined, the demonstrations become redundant.Motivated by this finding, we propose an efficient Progressive In-Context Alignment (PICA) method consisting of two stages.In the first few-shot stage, the model generates several prior response tokens via standard ICL while concurrently extracting the ICL vector that stores the task function from the separator token representation.In the following zero-shot stage, this ICL vector guides the model to generate responses without further demonstrations.Extensive experiments demonstrate that our PICA not only surpasses vanilla ICL but also achieves comparable performance to other alignment tuning methods.The proposed training-free method reduces the time cost (e.g., 5.45×) with improved alignment performance (e.g., 6.57+).Consequently, our work highlights the application of ICL for alignment and calls for a deeper understanding of ICL for complex generations.

Generally, the decoder-only large language models (LLMs) are adapted to context-aware neural machine translation (NMT) in a concatenating way, where LLMs take the concatenation of the source sentence (i.e., intrasentence context) and the inter-sentence context as the input, and then to generate the target tokens sequentially.This adaptation strategy, i.e., concatenation mode, considers intrasentence and inter-sentence contexts with the same priority, despite an apparent difference between the two kinds of contexts.In this paper, we propose an alternative adaptation approach, named Decoding-enhanced Multiphase Prompt Tuning (DeMPT), to make LLMs discriminately model and utilize the inter-and intra-sentence context and more effectively adapt LLMs to context-aware NMT.First, DeMPT divides the context-aware NMT process into three separate phases.During each phase, different continuous prompts are introduced to make LLMs discriminately model various information.Second, DeMPT employs a heuristic way to further discriminately enhance the utilization of the source-side interand intra-sentence information at the final decoding phase.Experiments show that our approach significantly outperforms the concatenation method, and further improves the performance of LLMs in discourse modeling.

With instruction tuning, Large Language Models (LLMs) can enhance their ability to adhere to commands.Diverging from most works focusing on data mixing, our study concentrates on enhancing the model's capabilities from the perspective of data sampling during training.Drawing inspiration from the human learning process, where it is generally easier to master solutions to similar topics through focused practice on a single type of topic, we introduce a novel instruction tuning strategy termed Com-monIT: Commonality-aware Instruction Tuning.Specifically, we cluster instruction datasets into distinct groups with three proposed metrics (TASK, EMBEDDING and LENGTH).We ensure each training mini-batch, or "partition", consists solely of data from a single group, which brings about both data randomness across mini-batches and intra-batch data similarity.Rigorous testing on LLaMa models demonstrates CommonIT's effectiveness in enhancing the instruction-following capabilities of LLMs through IT datasets (FLAN, CoT, and Alpaca) and models (LLaMa2-7B, Qwen2-7B, LLaMa 13B, and BLOOM 7B).CommonIT consistently boosts an average improvement of 2.1% on the general domain (i.e., the average score of Knowledge, Reasoning, Multilinguality and Coding) with the LENGTH metric, and 5.2% on the special domain (i.e., GSM, Openfunctions and Code) with the TASK metric, and 3.8% on the specific tasks (i.e., MMLU) with the EMBEDDING metric.Code is available at https://github.com/raojay7/CommonIT.

Text detoxification aims to minimize the risk of language models producing toxic content.However, existing detoxification methods fail to balance the detoxification effectiveness and generation quality.This issue arises from neglecting the constraints imposed by the context: language models are designed to generate output that closely matches the given context, while detoxification methods strive to ensure the safety of the output, even if it deviates semantically from the context.Given this, we introduce a Context-aware Model self-Detoxification (CMD) framework that pays attention to both the context and the detoxification process, i.e., first detoxifying the context and then making the language model generate along the safe context.Specifically, CMD framework involves two phases: utilizing language models to synthesize data and applying these data for training.We also introduce a toxic contrastive loss that encourages the model generation away from the negative toxic samples.Experiments on various LLMs have verified the effectiveness of our MSD framework, which can yield the best performance compared to baselines. 1 Warning: cases in this paper may contain offensive content.

Large language models (LLMs) exhibit remarkable performance across diverse tasks, indicating their potential for expansion into large speech-text models (LSMs) by integrating speech capabilities.Although unified speech-text pre-training and multimodal data instruction-tuning offer considerable benefits, these methods generally entail significant resource demands and tend to overfit specific tasks.This study aims to refine the use of speech datasets for LSM training by addressing the limitations of vanilla instruction tuning.We explore the instruction-following dynamics within LSMs, identifying a critical issue termed speech anchor bias-a tendency for LSMs to over-rely on speech inputs, mistakenly interpreting the entire speech modality as directives, thereby neglecting textual instructions.To counteract this bias, we introduce a self-powered LSM that leverages augmented automatic speech recognition data generated by the model itself for more effective instruction tuning.Our experiments across a range of speech-based tasks demonstrate that selfpowered LSM mitigates speech anchor bias and improves the fusion of speech and text modalities in LSMs.

Recent studies in Retrieval-Augmented Generation (RAG) have investigated extracting evidence from retrieved passages to reduce computational costs and enhance the final RAG performance, yet it remains challenging.Existing methods heavily rely on heuristic-based augmentation, encountering several issues: (1) Poor generalization due to hand-crafted context filtering; (2) Semantics deficiency due to rulebased context chunking; (3) Skewed length due to sentence-wise filter learning.To address these issues, we propose a model-based evidence extraction learning framework, SEER, optimizing a vanilla model as an evidence extractor with desired properties through selfaligned learning.Extensive experiments show that our method largely improves the final RAG performance, enhances the faithfulness, helpfulness, and conciseness of the extracted evidence, and reduces the evidence length by 9.25 times.The code will be available at https://github.com/HITsz-TMG/SEER.

In this paper, we study the problem of finding proper tradeoff for graph self-supervised learning. Recently, various self-supervised auxiliary tasks have been proposed to accelerate representation learning in Graph Neural Networks (GNNs). However, existing graph self-supervised learning ignores the task conflicts between the main task and auxiliary tasks. In this paper, we propose Pareto graph self-supervised learning, a general learning framework that not only finds the solution where the main task rather than auxiliary tasks achieves the best performance in graph self-supervised learning, but more importantly, learns the personalized self-supervised task for different nodes. The proposed method first formulates the graph self-supervised learning problem as a multi-objective optimization problem, and then solves the problem with the preferred vector and personalized optimization. Experimental results demonstrate the effectiveness of the proposed method by achieving state-of-the-art performance.

Identifying mild cognitive impairment (MCI) is vital for Alzheimer’s disease prevention. As neurodegenerative diseases progress, synchronous activity in electroencephalography (EEG) - indicating functional connectivity - changes due to neural system deterioration. Thus, developing geometric learning to decode the functional brain structure is essential. Techniques such as graph neural networks and Riemannian manifolds show potential in analyzing non-Euclidean data. However, existing approaches neglect to combine synchronous activity with temporal dependence and still remain insufficient for MCI detection. This paper proposes the Brain Network sequence-driven Manifold-based Transformer (BNMTrans) to identify MCI patterns from EEG data. BNMTrans leverages its strengths by extracting features from sequential brain networks through the self-attention mechanism, guided by the geometric correlations within the Riemannian manifold. By integrating long-term temporal dynamics and structural relationships within manifold space based on functional connectivity, this approach outperforms others in EEG feature comparisons and state-of-the-art evaluations based on clinical data from 89 subjects (46 MCI, 43 healthy controls) at a local hospital. Our work has significance for both MCI clinical management and technical progression in the EEG field.

Automatic depression detection based on audio and text representations from participants’ interviews has attracted widespread attention. However, most of previous researches only used one type of feature of one single modality for depression detection, so that the rich information of audio and text from interviews has not been fully utilized. Moreover, an effective multi-modal fusion approach to leverage the independence among audio and text representations is still lacking. To address these problems, we propose a multi-modal fusion depression detection model based on the interaction of multilevel audio features and text sentence embedding. Specifically, we first extract Low-Level Descriptors (LLDs), mel-spectrogram features, and wav2vec features from the audio. Then we design a Multi-level Audio Features Interaction Module (MAFIM) to fuse these three levels of features for a comprehensive audio representation. For interview text, we use pre-trained BERT to extract sentence-level embedding. Further, to effectively fuse audio and text representations, we design a Channel Attention-based Multi-modal Fusion Module (CAMFM) by taking into account the independence and correlation between two different modalities. Our proposed model shows better performance on two datasets, DAIC-WOZ and EATD-Corpus, than existing methods, so it has a high potential to be applied for interview-based depression detection in practice.

In nested Named entity recognition (NER), entities are nested with each other, and thus requiring more data annotations to address. This leads to the development of few-shot nested NER, where the prevalence of pretrained language models with in-context learning (ICL) offers promising solutions. In this work, we introduce an effective and innovative ICL framework for the setting of few-shot nested NER. We improve the ICL prompt by devising a novel example demonstration selection mechanism, EnDe retriever. In EnDe retriever, we employ contrastive learning to perform three types of representation learning, in terms of semantic similarity, boundary similarity, and label similarity, to generate high-quality demonstration examples. Extensive experiments over three nested NER and four flat NER datasets demonstrate the efficacy of our system.

Emotion recognition from electroencephalography (EEG) has attracted widespread interest, but few studies have considered estimating the highly dynamic trajectories of emotion in a relatively long period, such as video watching. To address this problem, we first recruit participants to assign continuous-time and continuous-level emotion labels to videos from the SEED corpus. Then, we propose a hybrid model, namely Emotion Time-Varying Regression (EmoTVR), to estimate continuous-time and continuous-level emotion using EEG spatial-temporal representations. EmoTVR combines atrous convolutional networks for spatial feature extraction and a temporal self-attentive regressor using attention-based long short-term memory for temporal feature extraction and continuous estimation. Moreover, EmoTVR adopts the Domain Adversarial Neural Network to address the problem of individual difference. Experimental results through both within-subject and cross-subject cross-validations demonstrate the superiority of EmoTVR in recognizing dynamic and continuous emotion over traditional methods. The proposed EmoTVR method caters for the needs of dynamic and continuous emotion recognition in naturalistic conditions, so it is highly potential for practical applications of emotion recognition.

Language modeling at scale has proven very effective and brought unprecedented success to natural language models. Many typical representatives, especially decoder-only models, e.g., BLOOM and LLaMA, and encoder-decoder models, e.g., Flan-T5 and AlexaTM, have exhibited incredible instruction-following capabilities while keeping strong task completion ability. These large language models can achieve superior performance in various tasks and even yield emergent capabilities, e.g., reasoning and universal generalization. Though the above two paradigms are mainstream and well explored, the potential of the BERT family, which are encoder-only based models and have ever been one of the most representative pre-trained models, also deserves attention, at least should be discussed. In this work, we adopt XML-R to explore the effectiveness of the BERT family for instruction following and zero-shot learning. We first design a simple yet effective strategy to utilize the encoder-only models for generation tasks and then conduct multi-task instruction tuning. Experimental results demonstrate that our fine-tuned model, Instruct-XMLR, outperforms Bloomz on all evaluation tasks and achieves comparable performance with mT0 on most tasks. Surprisingly, Instruct-XMLR also possesses strong task and language generalization abilities, indicating that Instruct-XMLR can also serve as a good instruction follower and zero-shot learner. Besides, Instruct-XMLR can accelerate decoding due to its non-autoregressive generation manner, achieving around 3 times speedup compared with current autoregressive large language models. Although we also witnessed several limitations through our experiments, such as the performance decline in long-generation tasks and the shortcoming of length prediction, Instruct-XMLR can still become a good member of the family of current large language models.

Text-based person search aims at retrieving images of a particular person based on a given textual description. A common solution for this task is to directly match the entire images and texts, i.e., global alignment, which fails to deal with discerning specific details that discriminate against appearance-similar people. As a result, some works shift their attention towards local alignment. One group matches fine-grained parts using forward attention weights of the transformer yet underutilizes information. Another implicitly conducts local alignment by reconstructing masked parts based on unmasked context yet with a biased masking strategy. All limit performance improvement. This paper proposes the Local Alignment from Image-Phrase modeling (LAIP) framework, with Bidirectional Attention-weighted local alignment (BidirAtt) and Mask Phrase Modeling (MPM) module. BidirAtt goes beyond the typical forward attention by considering the gradient of the transformer as backward attention, utilizing two-sided information for local alignment. MPM focuses on mask reconstruction within the noun phrase rather than the entire text, ensuring an unbiased masking strategy. Extensive experiments conducted on the CUHK-PEDES, ICFG-PEDES, and RSTPReid datasets demonstrate the superiority of the LAIP framework over existing methods.

Large Multimodal Models (LMMs) have achieved impressive success in visual reasoning, particularly in visual mathematics. However, problem-solving capabilities in graph theory remain less explored for LMMs, despite being a crucial aspect of mathematical reasoning that requires an accurate understanding of graphical structures and multi-step reasoning on visual graphs. To step forward in this direction, we are the first to design a benchmark named VisionGraph, used to explore the capabilities of advanced LMMs in solving multimodal graph theory problems. It encompasses eight complex graph problem tasks, from connectivity to shortest path problems. Subsequently, we present a Description-Program-Reasoning (DPR) chain to enhance the logical accuracy of reasoning processes through graphical structure description generation and algorithm-aware multi-step reasoning. Our extensive study shows that 1) GPT-4V outperforms Gemini Pro in multi-step graph reasoning; 2) All LMMs exhibit inferior perception accuracy for graphical structures, whether in zero/few-shot settings or with supervised fine-tuning (SFT), which further affects problem-solving performance; 3) DPR significantly improves the multi-step graph reasoning capabilities of LMMs and the GPT-4V (DPR) agent achieves SOTA performance.

Multi-party dialogue reading comprehension is an extraction-based reading comprehension task that aims to understand dialogue with multiple interlocutors and answer related questions. The frequent rotation of topics and the irregular order of interlocutors in dialogues may lead to the scattered distribution of information in multi-party dialogues. This means that the model needs to effectively integrate information across multiple utterances and among various interlocutors. Although previous methods have made considerable efforts in mining and modeling dialogue-related features, they still encounter two key issues. On the one hand, these methods failed to solve the cross-utterance co-reference problem that arises from the coexistence of multiple topics and interlocutors in the dialogue. On the other hand, they mostly ignored the joint reasoning of multi-granularity dialogue-related features, which can parse the semantic space of multi-party dialogue from coarse to fine. To overcome these bottlenecks, we propose a dual-granularity information joint reasoning method, which performs hierarchically semantic modeling for multi-party dialogue based on the graph attention networks. Specifically, we utilize discourse dependency relationships and interlocutor-aware temporal information to conduct coarsegrained semantic modeling, and perform fine-grained semantic refinement by leveraging token-level co-reference relationships. Our method demonstrates stable and substantial performance improvement when using different pre-trained language models as backbones and achieves a new state-of-the-art on the benchmark corpora Molweni and FriendsQA.

Cross-center data heterogeneity and annotation unreliability significantly challenge the intelligent diagnosis of diseases using brain signals. A notable example is the EEG-based diagnosis of neurodegenerative diseases, which features subtler abnormal neural dynamics typically observed in small-group settings. To advance this area, in this work, we introduce a transferable framework employing Manifold Attention and Confidence Stratification (MACS) to diagnose neurodegenerative disorders based on EEG signals sourced from four centers with unreliable annotations. The MACS framework's effectiveness stems from these features: 1) The Augmentor generates various EEG-represented brain variants to enrich the data space; 2) The Switcher enhances the feature space for trusted samples and reduces overfitting on incorrectly labeled samples; 3) The Encoder uses the Riemannian manifold and Euclidean metrics to capture spatiotemporal variations and dynamic synchronization in EEG; 4) The Projector, equipped with dual heads, monitors consistency across multiple brain variants and ensures diagnostic accuracy; 5) The Stratifier adaptively stratifies learned samples by confidence levels throughout the training process; 6) Forward and backpropagation in MACS are constrained by confidence stratification to stabilize the learning system amid unreliable annotations. Our subject-independent experiments, conducted on both neurocognitive and movement disorders using cross-center corpora, have demonstrated superior performance compared to existing related algorithms. This work not only improves EEG-based diagnostics for cross-center and small-setting brain diseases but also offers insights into extending MACS techniques to other data analyses, tackling data heterogeneity and annotation unreliability in multimedia and multimodal content understanding. We have released our code here: https://github.com/ICI-BCI/EEG-MACS.

Event extraction (EE) is a critical direction in the field of information extraction, laying an important foundation for the construction of structured knowledge bases. EE from text has received ample research and attention for years, yet there can be numerous real-world applications that require direct information acquisition from speech signals, online meeting minutes, interview summaries, press releases, etc. While EE from speech has remained under-explored, this paper fills the gap by pioneering a SpeechEE, defined as detecting the event predicates and arguments from a given audio speech. To benchmark the SpeechEE task, we first construct a large-scale high-quality dataset. Based on textual EE datasets under the sentence, document, and dialogue scenarios, we convert texts into speeches through both manual real-person narration and automatic synthesis, empowering the data with diverse scenarios, languages, domains, ambiences, and speaker styles. Further, to effectively address the key challenges in the task, we tailor an E2E SpeechEE system based on the encoder-decoder architecture, where a novel Shrinking Unit module and a retrieval-aided decoding mechanism are devised. Extensive experimental results on all SpeechEE subsets demonstrate the efficacy of the proposed model, offering a strong baseline for the task. At last, being the first work on this topic, we shed light on key directions for future research. Our codes and the benchmark datasets are open at https://SpeechEE.github.io/

The Visual Spatial Description Challenge (VSD) is the first competition event focused on visual spatial understanding, organized under the auspices of the ACM Multimedia Conference 2024. The goal of the VSD challenge is to assess the the ability of models and systems to comprehend spatial concepts, relationships and other semantics from a scene presented with visual appearance. The VSD challenge provides two benchmark datasets for three subtasks, i.e., visual spatial relationship classification, single spatial description generation, and open-ended spatial description generation. The challenge details are available on https://lllogen.github.io/vsd-challenge.github.io/.

While fine-tuning pretrained models has become common practice, these models often underperform outside their specific domains. Recently developed model merging techniques enable the direct integration of multiple models, each fine-tuned for distinct tasks, into a single model. This strategy promotes multitasking capabilities without requiring retraining on the original datasets. However, existing methods fall short in addressing potential conflicts and complex correlations between tasks, especially in parameter-level adjustments, posing a challenge in effectively balancing parameter competition across various tasks. This paper introduces an innovative technique named **PCB-Merging** (Parameter Competition Balancing), a *lightweight* and *training-free* technique that adjusts the coefficients of each parameter for effective model merging. PCB-Merging employs intra-balancing to gauge parameter significance within individual tasks and inter-balancing to assess parameter similarities across different tasks. Parameters with low importance scores are dropped, and the remaining ones are rescaled to form the final merged model. We assessed our approach in diverse merging scenarios, including cross-task, cross-domain, and cross-training configurations, as well as out-of-domain generalization. The experimental results reveal that our approach achieves substantial performance enhancements across multiple modalities, domains, model sizes, number of tasks, fine-tuning forms, and large language models, outperforming existing model merging methods.

Despite their remarkable abilities in various tasks, large language models (LLMs) still struggle with real-time information (e.g., new facts and terms) due to the knowledge cutoff in their development process. However, existing benchmarks focus on outdated content and limited fields, facing difficulties in real-time updating and leaving new terms unexplored. To address this problem, we propose an adaptive benchmark, NewTerm, for real-time evaluation of new terms. We design a highly automated construction method to ensure high-quality benchmark construction with minimal human effort, allowing flexible updates for real-time information. Empirical results on various LLMs demonstrate over 20% performance reduction caused by new terms. Additionally, while updates to the knowledge cutoff of LLMs can cover some of the new terms, they are unable to generalize to more distant new terms. We also analyze which types of terms are more challenging and why LLMs struggle with new terms, paving the way for future research. Finally, we construct NewTerm 2022 and 2023 to evaluate the new terms updated each year and will continue updating annually. The benchmark and codes can be found at https://anonymous.4open.science/r/NewTerms.

Large Language Models (LLMs) have exhibited an impressive ability to perform In-Context Learning (ICL) from only a few examples. Recent works have indicated that the functions learned by ICL can be represented through compressed vectors derived from the transformer. However, the working mechanisms and optimization of these vectors are yet to be thoroughly explored. In this paper, we address this gap by presenting a comprehensive analysis of these compressed vectors, drawing parallels to the parameters trained with gradient descent, and introducing the concept of state vector. Inspired by the works on model soup and momentum-based gradient descent, we propose inner and momentum optimization methods that are applied to refine the state vector progressively as test-time adaptation. Moreover, we simulate state vector aggregation in the multiple example setting, where demonstrations comprising numerous examples are usually too lengthy for regular ICL, and further propose a divide-and-conquer aggregation method to address this challenge. We conduct extensive experiments using Llama-2 and GPT-J in both zero-shot setting and few-shot setting. The experimental results show that our optimization method effectively enhances the state vector and achieves the state-of-the-art performance on diverse tasks.

Instruction tuning (IT) is crucial to tailoring large language models (LLMs) towards human-centric interactions. Recent advancements have shown that the careful selection of a small, high-quality subset of IT data can significantly enhance the performance of LLMs. Despite this, common approaches often rely on additional models or data, which increases costs and limits widespread adoption. In this work, we propose a novel approach, termed $\textit{SelectIT}$, that capitalizes on the foundational capabilities of the LLM itself. Specifically, we exploit the intrinsic uncertainty present in LLMs to more effectively select high-quality IT data, without the need for extra resources. Furthermore, we introduce a curated IT dataset, the $\textit{Selective Alpaca}$, created by applying SelectIT to the Alpaca-GPT4 dataset. Empirical results demonstrate that IT using Selective Alpaca leads to substantial model ability enhancement. The robustness of SelectIT has also been corroborated in various foundation models and domain-specific tasks. Our findings suggest that longer and more computationally intensive IT data may serve as superior sources of IT, offering valuable insights for future research in this area. Data, code, and scripts are freely available at https://github.com/Blue-Raincoat/SelectIT.

This paper introduces AutoSurvey, a speedy and well-organized methodology for automating the creation of comprehensive literature surveys in rapidly evolving fields like artificial intelligence. Traditional survey paper creation faces challenges due to the vast volume and complexity of information, prompting the need for efficient survey methods. While large language models (LLMs) offer promise in automating this process, challenges such as context window limitations, parametric knowledge constraints, and the lack of evaluation benchmarks remain. AutoSurvey addresses these challenges through a systematic approach that involves initial retrieval and outline generation, subsection drafting by specialized LLMs, integration and refinement, and rigorous evaluation and iteration. Our contributions include a comprehensive solution to the survey problem, a reliable evaluation method, and experimental validation demonstrating AutoSurvey's effectiveness.

This work focuses on the active learning in video captioning. In particular, we propose to address the learnability problem in active learning, which has been brought up by collective outliers in video captioning and neglected in the literature. To start with, we conduct a comprehensive study of collective outliers, exploring their hard-to-learn property and concluding that ground truth inconsistency is one of the main causes. Motivated by this, we design a novel active learning algorithm that takes three complementary aspects, namely learnability, diversity, and uncertainty, into account. Ideally, learnability is reflected by ground truth consistency. Under the active learning scenario where ground truths are not available until human involvement, we measure the consistency on estimated ground truths, where predictions from off-the-shelf models are utilized as approximations to ground truths. These predictions are further used to estimate sample frequency and reliability, evincing the diversity and uncertainty respectively. With the help of our novel caption-wise active learning protocol, our algorithm is capable of leveraging knowledge from humans in a more effective yet intellectual manner. Results on publicly available video captioning datasets with diverse video captioning models demonstrate that our algorithm outperforms SOTA active learning methods by a large margin, e.g. we achieve about 103% of full performance on CIDEr with 25% of human annotations on MSR-VTT.

In the visual spatial understanding (VSU) field, spatial image-to-text (SI2T) and spatial text-to-image (ST2I) are two fundamental tasks that appear in dual form. Existing methods for standalone SI2T or ST2I perform imperfectly in spatial understanding, due to the difficulty of 3D-wise spatial feature modeling. In this work, we consider modeling the SI2T and ST2I together under a dual learning framework. During the dual framework, we then propose to represent the 3D spatial scene features with a novel 3D scene graph (3DSG) representation that can be shared and beneficial to both tasks. Further, inspired by the intuition that the easier 3D$\to$image and 3D$\to$text processes also exist symmetrically in the ST2I and SI2T, respectively, we propose the Spatial Dual Discrete Diffusion (SD$^3$) framework, which utilizes the intermediate features of the 3D$\to$X processes to guide the hard X$\to$3D processes, such that the overall ST2I and SI2T will benefit each other. On the visual spatial understanding dataset VSD, our system outperforms the mainstream T2I and I2T methods significantly. Further in-depth analysis reveals how our dual learning strategy advances.

demands substantial human effort and incurs high training

Misspellings in search engine queries may prevent search engines from returning accurate results. For Chinese mobile search engines, due to the different input methods (e.g., hand-written and T9 input methods), more types of misspellings exist, making this problem more challenging. As an essential module of search engines, Chinese Spelling Check~(CSC) models aim to detect and correct misspelled Chinese characters from user-issued queries. Despite the great value of CSC to the search engine, there is no CSC benchmark collected from real-world search engine queries. To fill this blank, we construct and release the Alipay Search Engine Query (AlipaySEQ) spelling check dataset. To the best of our knowledge, AlipaySEQ is the first Chinese Spelling Check dataset collected from the real-world scenario of Chinese mobile search engines. It consists of 15,522 high-quality human annotated and 1,175,151 automatically generated samples. To demonstrate the unique challenges of AlipaySEQ in the era of Large Language Models~(LLMs), we conduct a thorough study to analyze the difference between AlipaySEQ and existing SIGHAN benchmarks and compare the performance of various baselines, including existing task-specific methods and LLMs. We observe that all baselines fail to perform satisfactorily due to the over-correction problem. Especially, LLMs exhibit below-par performance on AlipaySEQ, which is rather surprising. Therefore, to alleviate the over-correction problem, we introduce a model-agnostic CSC Self-Refine Framework (SRF) to construct a strong baseline. Comprehensive experiments demonstrate that our proposed SRF, though more effective against existing models on both the AlipaySEQ and SIGHAN15, is still far from achieving satisfactory performance on our real-world dataset. With the newly collected real-world dataset and strong baseline, we hope more progress can be achieved on such a challenging and valuable task.

Abstract Dialogue-level dependency parsing, despite its growing academic interest, often encounters underperformance issues due to resource shortages. A potential solution to this challenge is data augmentation. In recent years, large language models (LLMs) have demonstrated strong capabilities in generation, which can facilitate data augmentation greatly. In this study, we focus on Chinese dialogue-level dependency parsing, presenting three simple and effective strategies with LLM to augment the original training instances, namely word-level, syntax-level, and discourse-level augmentations, respectively. These strategies enable LLMs to either preserve or modify dependency structures, thereby assuring accuracy while increasing the diversity of instances at different levels. We conduct experiments on the benchmark dataset released by Jiang et al. (2023) to validate our approach. Results show that our method can greatly boost the parsing performance in various settings, particularly in dependencies among elementary discourse units. Lastly, we provide in-depth analysis to show the key points of our data augmentation strategies.

Deploying compute-intensive deep neural networks (DNNs) on resource-constrained end devices has become a prominent trend, enabling localized intelligence. However, efficiently deploying these DNNs at scale poses challenges. To address this, extensive research has focused on the early exit architecture based on convolutional neural networks (CNNs), which dynamically adapt network depth to reduce inference computation. Nevertheless, the sequential execution of all internal classifiers (ICs) and subsequent termination based on an exit criterion is inefficient. Motivated by these insights, we introduce a resolution-adaptive prediction network (RAPNet) architecture. RAPNet comprises a lightweight prediction network that captures global image features and an inference network integrated with an early exit architecture. The prediction network accurately determines the optimal IC position conditioned on the input images for efficient image classification. Additionally, we incorporate resolution-adaptive inference and feature fusion mechanisms by computational reuse, to effectively mitigate image spatial redundancy and improve the accuracy of ICs. We conduct extensive experiments across various data sets and architectures to demonstrate that RAPNet achieves a significantly better accuracy versus computational tradeoff than other recently proposed early exit methods. For instance, when using MobileNet as the base network, RAPNet achieves significant accuracy improvements of 12% and 5.7% on the Tiny Imagenet and CIFAR-100 data sets, respectively, surpassing other early exit methods with similar computational constraints.

Knowledge transfer is investigated in various natural language processing tasks except cross-domain constituency parsing. In this paper, we leverage heterogeneous data to transfer cross-domain and cross-task knowledge to constituency parsing. Concretely, we first select language modeling, named entity recognition, CCG supertagging and dependency parsing as auxiliary tasks and collect the corpora of these tasks covering various domains as cross-domain and cross-task heterogeneous data. Second, we exploit three types of prefixes: shared, task and domain prefix, to merge cross-domain and cross-task data and decompose the general, task and domain representation in the pretrained language model. Third, we convert the data formats of multi-source heterogeneous datasets and loss objectives of the auxiliary tasks into a consistent formalization closer to constituency parsing. Finally, we jointly train the model to transfer task and domain knowledge to cross-domain constituency parsing. We verify the effectiveness of our proposed model on five target domains of MCTB. Experimental results show that our knowledge transfer model outperforms various baseline models, including conventional chart-based and transition-based parsers and the current large-scale language model for zero-shot and few-shot settings.

While pre-training large-scale video-language models (VLMs) has shown remarkable potential for various downstream video-language tasks, existing VLMs can still suffer from certain commonly seen limitations, e.g., coarse-grained cross-modal aligning, under-modeling of temporal dynamics, detached video-language view. In this work, we target enhancing VLMs with a fine-grained structural spatio-temporal alignment learning method (namely Finsta). First of all, we represent the input texts and videos with fine-grained scene graph (SG) structures, both of which are further unified into a holistic SG (HSG) for bridging two modalities. Then, an SG-based framework is built, where the textual SG (TSG) is encoded with a graph Transformer, while the video dynamic SG (DSG) and the HSG are modeled with a novel recurrent graph Transformer for spatial and temporal feature propagation. A spatial-temporal Gaussian differential graph Transformer is further devised to strengthen the sense of the changes in objects across spatial and temporal dimensions. Next, based on the fine-grained structural features of TSG and DSG, we perform object-centered spatial alignment and predicate-centered temporal alignment respectively, enhancing the video-language grounding in both the spatiality and temporality. We design our method as a plug&play system, which can be integrated into existing well-trained VLMs for further representation augmentation, without training from scratch or relying on SG annotations in downstream applications. On 6 representative VL modeling tasks over 12 datasets in both standard and long-form video scenarios, Finsta consistently improves the existing 13 strong-performing VLMs persistently, and refreshes the current state-of-the-art end task performance significantly in both the fine-tuning and zero-shot settings.

Query-oriented micro-video summarization task aims to generate a concise sentence with two properties: (a) summarizing the main semantic of the micro-video and (b) being expressed in the form of search queries to facilitate retrieval. Despite its enormous application value in the retrieval area, this direction has barely been explored. Previous studies of summarization mostly focus on the content summarization for traditional long videos. Directly applying these studies is prone to gain unsatisfactory results because of the unique features of micro-videos and queries: diverse entities and complex scenes within a short time, semantic gaps between modalities, and various queries in distinct expressions. To specifically adapt to these characteristics, we propose a query-oriented micro-video summarization model, dubbed QMS. It employs an encoder-decoder-based transformer architecture as the skeleton. The multi-modal (visual and textual) signals are passed through two modal-specific encoders to obtain their representations, followed by an entity-aware representation learning module to identify and highlight critical entity information. As to the optimization, regarding the large semantic gaps between modalities, we assign different confidence scores according to their semantic relevance in the optimization process. Additionally, we develop a novel strategy to sample the effective target query among the diverse query set with various expressions. Extensive experiments demonstrate the superiority of the QMS scheme, on both the summarization and retrieval tasks, over several state-of-the-art methods.

Randomness is widely introduced in neural network training to simplify model optimization or avoid the over-fitting problem. Among them, dropout and its variations in different aspects (e.g., data, model structure) are prevalent in regularizing the training of deep neural networks. Though effective and performing well, the randomness introduced by these dropout-based methods causes nonnegligible inconsistency between training and inference. In this paper, we introduce a simple consistency training strategy to regularize such randomness, namely R-Drop, which forces two output distributions sampled by each type of randomness to be consistent. Specifically, R-Drop minimizes the bidirectional KL-divergence between two output distributions produced by dropout-based randomness for each training sample. Theoretical analysis reveals that R-Drop can reduce the above inconsistency by reducing the inconsistency among the sampled sub structures and bridging the gap between the loss calculated by the full model and sub structures. Experiments on 7 widely-used deep learning tasks ( 23 datasets in total) demonstrate that R-Drop is universally effective for different types of neural networks (i.e., feed-forward, recurrent, and graph neural networks) and different learning paradigms (supervised, parameter-efficient, and semi-supervised). In particular, it achieves state-of-the-art performances with the vanilla Transformer model on WMT14 English → German translation ( 30.91 BLEU) and WMT14 English → French translation ( 43.95 BLEU), even surpassing models trained with extra large-scale data and expert-designed advanced variants of Transformer models.

Abstract Hierarchical text classification (HTC) is an important task with broad applications, and few-shot HTC has gained increasing interest recently. While in-context learning (ICL) with large language models (LLMs) has achieved significant success in few-shot learning, it is not as effective for HTC because of the expansive hierarchical label sets and extremely ambiguous labels. In this work, we introduce the first ICL-based framework with LLM for few-shot HTC. We exploit a retrieval database to identify relevant demonstrations, and an iterative policy to manage multi-layer hierarchical labels. Particularly, we equip the retrieval database with HTC label-aware representations for the input texts, which is achieved by continual training on a pretrained language model with masked language modeling (MLM), layer-wise classification (CLS, specifically for HTC), and a novel divergent contrastive learning (DCL, mainly for adjacent semantically similar labels) objective. Experimental results on three benchmark datasets demonstrate superior performance of our method, and we can achieve state-of-the-art results in few-shot HTC.

Recognizing human emotions from complex, multivariate, and non-stationary electroencephalography (EEG) time series is essential in affective brain-computer interface. However, because continuous labeling of ever-changing emotional states is not feasible in practice, existing methods can only assign a fixed label to all EEG timepoints in a continuous emotion-evoking trial, which overlooks the highly dynamic emotional states and highly non-stationary EEG signals. To solve the problems of high reliance on fixed labels and ignorance of time-changing information, in this paper we propose a time-aware sampling network (TAS-Net) using deep reinforcement learning (DRL) for unsupervised emotion recognition, which is able to detect key emotion fragments and disregard irrelevant and misleading parts. Specifically, we formulate the process of mining key emotion fragments from EEG time series as a Markov decision process and train a time-aware agent through DRL without label information. First, the time-aware agent takes deep features from a feature extractor as input and generates sample-wise importance scores reflecting the emotion-related information each sample contains. Then, based on the obtained sample-wise importance scores, our method preserves top-Xcontinuous EEG fragments with relevant emotion and discards the rest. Finally, we treat these continuous fragments as key emotion fragments and feed them into a hypergraph decoding model for unsupervised clustering. Extensive experiments are conducted on three public datasets (SEED, DEAP, and MAHNOB-HCI) for emotion recognition using leave-one-subject-out cross-validation, and the results demonstrate the superiority of the proposed method against previous unsupervised emotion recognition methods. The proposed TAS-Net has great potential in achieving a more practical and accurate affective brain-computer interface in a dynamic and label-free circumstance. The source code is made available athttps://github.com/infinite-tao/TAS-Net.

Label noise is an important issue in machine learning, which might lead to negative influences on various tasks. Given that real benchmarks for evaluation of noise reduction methods are limited, plenty of studies construct pseudo noisy data to verify their proposed methods. However, very few works have realized the rationality of the noise generation strategies. If the generated pseudo datasets are biased, their final conclusions might also be problematic. In this work, we focus on text classification of natural language processing (NLP) to investigate various pseudo noise generation methods, which is the first work of this line for NLP. In particular, we compare the noise generated with crowdsourcing noise, a kind of real noise as gold-standard, to evaluate these noise generation methods. After then, we measure and compare the performance of representative noise reduction methods respectively based on the data of crowdsourcing and our top-ranked pseudo noisy generation strategies. We conduct experiments on five text classification datasets, offering detailed comparison results as well as discussions.

Natural language processing (NLP) has recently shown significant progress in rich-resource scenarios. However, it is much less effective for low-resource scenarios due to the model easily overfitting to limited training data and generalizing poorly on testing data. In recent years, consistency training has been widely adopted and shown great promise in deep learning, but still remains unexplored in low-resource settings. In this work, we propose DM-CT, a framework that incorporates both data-level and model-level consistency training as well as advanced data augmentation techniques for low-resource scenarios. Concretely, the input data is first augmented, and the output distributions of different sub-models generated by model variance are forced to be consistent (model-level consistency). Meanwhile, the predictions of the original input and the augmented one are also constrained to be consistent (data-level consistency). Experiments on different low-resource NLP tasks, including neural machine translation (4 IWSLT14 translation tasks, multilingual translation task, and WMT16 Romanian$\to$English translation), natural language understanding tasks (GLUE benchmark), and named entity recognition (Conll2003 and WikiGold), well demonstrate the superiority of DM-CT by obtaining significant and consistent performance improvements.

Image-text retrieval is a fundamental task to model a connection between images and natural language. Under its flourishing development in performance, most current methods suffer fromN-related time complexity, which hinders their application in practice to a certain extent. Targeting efficiency improvement, we propose a simple and effective keyword-guided pre-screening framework for image-text retrieval. Specifically, we convert the image and text data into keywords and perform keyword matching across the modalities to exclude a large number of irrelevant gallery samples prior to the retrieval network. For the keyword prediction, we transfer it into a multi-label classification problem and propose a multi-task learning scheme by appending the multi-label classifiers to the image-text retrieval network to achieve a lightweight and high-performance keyword prediction. For keyword matching, we introduce the inverted index from the search engine and thus create a win-win situation on both time and space complexities for the pre-screening. Extensive experiments on the two widely-used datasets,i.e., Flickr30K and MS-COCO, verify the effectiveness of the proposed framework. The proposed framework equipped with only two embedding layers achievesO(1) querying time complexity, while improving the retrieval efficiency and maintaining performance, when applied prior to the common image-text retrieval methods.

Current talking face generation methods have achieved promising lip-synchronization results, while still struggling to generate talking face video that exhibits emotional expressions and head poses. Studies in psychology have demonstrated that people may manifest diverse facial animations that follow a time-varying distribution. This presents two stochastic challenges that make generating appropriate emotional expressions and head poses difficult: (1) modelling the time-varying distribution of facial deformations to synthesize the stochastic dynamics of emotional expressions and head poses, and (2) estimating the complex motion distribution with given audio features to capture ambiguous audio-related expressions and head poses. To address the above issues, we present a Stochastic Latent talkIng face Generation mOdel (SLIGO), which builds a deep state space model (SSM) for talking face generation. The SLIGO model captures diverse and stochastic facial dynamics via the latent motion distribution. Additionally, we devise a dynamic variational autoencoder (DVAE) method to optimize the deep SSM model. This method decomposes the Evidence Lower BOund (ELBO) of SSM into three components: a posterior for latent motion encoding, a prior for audio-driven motion prediction, and a likelihood for talking face decoding. Furthermore, we propose a novel mixer continuous normalizing flow (CNF) module to model the complex facial motion prior distribution. Experimental results demonstrate that SLIGO outperforms existing methods and achieves state-of-the-art performance.

The increasing deployment of multiple deep neural networks (DNNs) on edge devices is revolutionizing mobile vision applications, spanning autonomous vehicles, augmented reality, and video surveillance. These applications demand adaptation to contextual and environmental drifts, typically through fine-tuning on edge devices without cloud access, due to increasing data privacy concerns and the urgency for timely responses. However, fine-tuning multiple DNNs on edge devices faces significant challenges due to the substantial computational workload. In this paper, we present PatchLine, a novel framework tailored for efficient on-device training in the form of fine-tuning for multi-DNN vision applications. At the core of PatchLine is an innovative lightweight adapter design called patches coupled with a strategic patch updating approach across models. Specifically, PatchLine adopts drift-adaptive incremental patching, correlation-aware warm patching, and entropy-based sample selection, to holistically reduce the number of trainable parameters, training epochs, and training samples. Experiments on four datasets, three vision tasks, four backbones, and two platforms demonstrate that PatchLine reduces the total computational cost by an average of 55% without sacrificing accuracy compared to the state-of-the-art.

Current efficient approaches to building Multimodal Large Language Models (MLLMs) mainly incorporate visual information into LLMs with a simple visual mapping network such as a linear projection layer, a multilayer perceptron (MLP), or Q-former from BLIP-2. Such networks project the image feature once and do not consider the interaction between the image and the human inputs. Hence, the obtained visual information without being connected to human intention may be inadequate for LLMs to generate intention-following responses, which we refer to as static visual information. To alleviate this issue, our paper introduces LMEye, a human-like eye with a play-and-plug interactive perception network, designed to enable dynamic interaction between LLMs and external visual information. It can allow the LLM to request the desired visual information aligned with various human instructions, which we term dynamic visual information acquisition. Specifically, LMEye consists of a simple visual mapping network to provide the basic perception of an image for LLMs. It also contains additional modules responsible for acquiring requests from LLMs, performing request-based visual information seeking, and transmitting the resulting interacted visual information to LLMs, respectively. In this way, LLMs act to understand the human query, deliver the corresponding request to the request-based visual information interaction module, and generate the response based on the interleaved multimodal information. We evaluate LMEye through extensive experiments on multimodal benchmarks, demonstrating that it significantly improves zero-shot performances on various multimodal tasks compared to previous methods, with fewer parameters. Moreover, we also verify its effectiveness and scalability on various language models and video understanding, respectively.

Pre-trained models are frequently employed in multimodal learning. However, these models have too many parameters and need too much effort to fine-tune the downstream tasks. Knowledge distillation (KD) is a method to transfer knowledge using the soft label from this pre-trained teacher model to a smaller student, where the parameters of the teacher are fixed (or partially) during training. Recent studies show that this mode may cause difficulties in knowledge transfer due to the mismatched model capacities. To alleviate the mismatch problem, adjustment of temperature parameters, label smoothing and teacher-student joint training methods (online distillation) to smooth the soft label of a teacher network, have been proposed. But those methods rarely explain the effect of smoothed soft labels to enhance the KD performance. The main contributions of our work are the discovery, analysis, and validation of the effect of the smoothed soft label and a less time-consuming and adaptive transfer of the pre-trained teacher's knowledge method, namely PESF-KD by adaptive tuning soft labels of the teacher network. Technically, we first mathematically formulate the mismatch as the sharpness gap between teacher's and student's predictive distributions, where we show such a gap can be narrowed with the appropriate smoothness of the soft label. Then, we introduce an adapter module for the teacher and only update the adapter to obtain soft labels with appropriate smoothness. Experiments on various benchmarks including CV and NLP show that PESF-KD can significantly reduce the training cost while obtaining competitive results compared to advanced online distillation methods.

Simultaneous machine translation (SiMT) is usually done via sequence-level knowledge distillation (Seq-KD) from a full-sentence neural machine translation (NMT) model. However, there is still a significant performance gap between NMT and SiMT. In this work, we propose to leverage monolingual data to improve SiMT, which trains a SiMT student on the combination of bilingual data and external monolingual data distilled by Seq-KD. Preliminary experiments on En-Zh and En-Ja news domain corpora demonstrate that monolingual data can significantly improve translation quality (e.g., +3.15 BLEU on En-Zh). Inspired by the behavior of human simultaneous interpreters, we propose a novel monolingual sampling strategy for SiMT, considering both chunk length and monotonicity. Experimental results show that our sampling strategy consistently outperforms the random sampling strategy (and other conventional typical NMT monolingual sampling strategies) by avoiding the key problem of SiMT -- hallucination, and has better scalability. We achieve +0.72 BLEU improvements on average against random sampling on En-Zh and En-Ja. Data and codes can be found at https://github.com/hexuandeng/Mono4SiMT.

Non-autoregressive neural machine translation (NAT) models are proposed to accelerate the inference process while maintaining relatively high performance. However, existing NAT models are difficult to achieve the desired efficiency-quality trade-off. For one thing, fully NAT models with efficient inference perform inferior to their autoregressive counterparts. For another, iterative NAT models can, though, achieve comparable performance while diminishing the advantage of speed. In this paper, we propose RenewNAT, a flexible framework with high efficiency and effectiveness, to incorporate the merits of fully and iterative NAT models. RenewNAT first generates the potential translation results and then renews them in a single pass. It can achieve significant performance improvements at the same expense as traditional NAT models (without introducing additional model parameters and decoding latency). Experimental results on various translation benchmarks (e.g., 4 WMT) show that our framework consistently improves the performance of strong fully NAT methods (e.g., GLAT and DSLP) without additional speed overhead.

Deep neural models (e.g. Transformer) naturally learn spurious features, which create a ``shortcut'' between the labels and inputs, thus impairing the generalization and robustness. This paper advances self-attention mechanism to its robust variant for Transformer-based pre-trained language models (e.g. BERT). We propose Adversarial Self-Attention mechanism (ASA), which adversarially biases the attentions to effectively suppress the model reliance on features (e.g. specific keywords) and encourage its exploration of broader semantics. We conduct comprehensive evaluation across a wide range of tasks for both pre-training and fine-tuning stages. For pre-training, ASA unfolds remarkable performance gain compared to naive training for longer steps. For fine-tuning, ASA-empowered models outweigh naive models by a large margin considering both generalization and robustness.

Transformer-based autoregressive (AR) methods have achieved appealing performance for varied sequence-to-sequence generation tasks, e.g., neural machine translation, summarization, and code generation, but suffer from low inference efficiency. To speed up the inference stage, many non-autoregressive (NAR) strategies have been proposed in the past few years. Among them, the conditional masked language model (CMLM) is one of the most versatile frameworks, as it can support many different sequence generation scenarios and achieve very competitive performance on these tasks. In this paper, we further introduce a simple yet effective adaptive masking over masking strategy to enhance the refinement capability of the decoder and make the encoder optimization easier. Experiments on 3 different tasks (neural machine translation, summarization, and code generation) with 15 datasets in total confirm that our proposed simple method achieves significant performance improvement over the strong CMLM model. Surprisingly, our proposed model yields state-of-the-art performance on neural machine translation (34.62 BLEU on WMT16 EN to RO, 34.82 BLEU on WMT16 RO to EN, and 34.84 BLEU on IWSLT De to En) and even better performance than the AR Transformer on 7 benchmark datasets with at least 2.2x speedup. Our code is available at GitHub.

In this work, we investigate a more realistic unsupervised multimodal machine translation (UMMT) setup, inference-time image-free UMMT, where the model is trained with sourcetext image pairs, and tested with only sourcetext inputs.First, we represent the input images and texts with the visual and language scene graphs (SG), where such fine-grained visionlanguage features ensure a holistic understanding of the semantics.To enable pure-text input during inference, we devise a visual scene hallucination mechanism that dynamically generates pseudo visual SG from the given textual SG.Several SG-pivoting based learning objectives are introduced for unsupervised translation training.On the benchmark Multi30K data, our SG-based method outperforms the best-performing baseline by significant BLEU scores on the task and setup, helping yield translations with better completeness, relevance and fluency without relying on paired images.Further in-depth analyses reveal how our model advances in the task setting.

The ability of commonsense reasoning (CR) decides whether a neural machine translation (NMT) model can move beyond pattern recognition.Despite the rapid advancement of NMT and the use of pretraining to enhance NMT models, research on CR in NMT is still in its infancy, leaving much to be explored in terms of effectively training NMT models with high CR abilities and devising accurate automatic evaluation metrics.This paper presents a comprehensive study aimed at expanding the understanding of CR in NMT.For the training, we confirm the effectiveness of incorporating pretrained knowledge into NMT models and subsequently utilizing these models as robust testbeds for investigating CR in NMT.For the evaluation, we propose a novel entity-aware evaluation method that takes into account both the NMT candidate and important entities in the candidate, which is more aligned with human judgement.Based on the strong testbed and evaluation methods, we identify challenges in training NMT models with high CR abilities and suggest directions for further unlabeled data utilization and model design.We hope that our methods and findings will contribute to advancing the research of CR in NMT.

Zhiwei Cao, Baosong Yang, Huan Lin, Suhang Wu, Xiangpeng Wei, Dayiheng Liu, Jun Xie, Min Zhang, Jinsong Su. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Conditional inference on joint textual and visual clues is a multi-modal reasoning task that textual clues provide prior permutation or external knowledge, which are complementary with visual content and pivotal to deducing the correct option.Previous methods utilizing pretrained vision-language models (VLMs) have achieved impressive performances, yet they show a lack of multimodal context reasoning capability, especially for text-modal information.To address this issue, we propose a Multi-modal Context Reasoning approach, named ModCR.Compared to VLMs performing reasoning via cross modal semantic alignment, it regards the given textual abstract semantic and objective image information as the pre-context information and embeds them into the language model to perform context reasoning.Different from recent vision-aided language models used in natural language processing, ModCR incorporates the multi-view semantic alignment information between language and vision by introducing the learnable alignment prefix between image and text in the pretrained language model.This makes the language model well-suitable for such multi-modal reasoning scenario on joint textual and visual clues.We conduct extensive experiments on two corresponding data sets and experimental results show significantly improved performance (exact gain by 4.8% on PMR test set) compared to previous strong baselines.Code

Pretrained Vision-Language Models (VLMs) have achieved remarkable performance in image retrieval from text.However, their performance drops drastically when confronted with linguistically complex texts that they struggle to comprehend.Inspired by the Divide-and-Conquer (Smith, 1985) algorithm and dualprocess theory (Groves and Thompson, 1970), in this paper, we regard linguistically complex texts as compound proposition texts composed of multiple simple proposition sentences and propose an end-to-end Neural Divide-and-Conquer Reasoning framework, dubbed NDCR.It contains three main components: 1) Divide: a proposition generator divides the compound proposition text into simple proposition sentences and produces their corresponding representations, 2) Conquer: a pretrained VLMsbased visual-linguistic interactor achieves the interaction between decomposed proposition sentences and images, 3) Combine: a neuralsymbolic reasoner combines the above reasoning states to obtain the final solution via a neural logic reasoning approach.According to the dual-process theory, the visual-linguistic interactor and neural-symbolic reasoner could be regarded as analogical reasoning System 1 and logical reasoning System 2. We conduct extensive experiments on a challenging image retrieval from contextual descriptions data set.Experimental results and analyses indicate NDCR significantly improves performance in the complex image-text reasoning problem.

Yinghao Li, Xuebo Liu, Shuo Wang, Peiyuan Gong, Derek F. Wong, Yang Gao, Heyan Huang, Min Zhang. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Despite the excellent performance of Pretrained Language Models on many text generation tasks, they suffer from inefficient inference on computation and memory due to their largescale parameters and the universal autoregressive decoding paradigm.In this work, we propose a novel fine-tuning method DEER, which can make a single pre-trained model support Dynamic and Efficient infERence and achieve an adaptive trade-off between model performance and latency.In particular, our critical insight is to jointly utilize the non-autoregressive (NAR) generation and dynamic parameter pruning techniques, which can flexibly control the decoding iteration steps and model sizes according to memory and latency limitations.Besides, we also explore the effectiveness of the pre-trained MLMs (i.e., the BERT family) for text generation tasks since their bidirectional attention nature is more suitable for the NAR training objective.Extensive experiments on both monolingual and multilingual pre-trained MLMs demonstrate the effectiveness of our proposed DEER method by consistently achieving (1) higher BLEU scores than the strong autoregressive Transformer model on three neural machine translation tasks with 3 → 12 times speedup, (2) competitive performance (but with much faster inference speed) compared with the BART model on four GLGE benchmark tasks.Our code will be publicly available at GitHub 1 .

Pre-trained autoregressive (AR) language models such as BART and GPTs have dominated Open-ended Long Text Generation (Open-LTG).However, the AR nature will decrease the inference efficiency along with the increase of generation length, which hinder their application in Open-LTG.To improve inference efficiency, we alternatively explore the potential of the pre-trained masked language models (MLMs) along with a representative iterative non-autoregressive (NAR) decoding strategy for Open-LTG.Our preliminary study shows that pre-trained MLMs can merely generate short text and will collapse for long text modeling.To enhance the long text generation capability of MLMs, we introduce two simple yet effective strategies for the iterative NAR model: dynamic sliding window attention (DSWA) and linear temperature decay (LTD).It can alleviate long-distance collapse problems and achieve longer text generation with a flexible trade-off between performance and inference speedup.Experiments on the storytelling and multi-paragraph opinionated article writing tasks show that pre-trained MLMs can achieve more than 3 × → 13 × speedup with better performance than strong AR models.Our code is available at GitHub * .

Shudong Liu, Xuebo Liu, Derek F. Wong, Zhaocong Li, Wenxiang Jiao, Lidia S. Chao, Min Zhang. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

The neural metrics recently received considerable attention from the research community in the automatic evaluation of machine translation.Unlike text-based metrics that have interpretable and consistent evaluation mechanisms for various data sources, the reliability of neural metrics in assessing out-of-distribution data remains a concern due to the disparity between training data and real-world data.This paper aims to address the inference bias of neural metrics through uncertainty minimization during test time, without requiring additional data.Our proposed method comprises three steps: uncertainty estimation, test-time adaptation, and inference.Specifically, the model employs the prediction uncertainty of the current data as a signal to update a small fraction of parameters during test time and subsequently refine the prediction through optimization.To validate our approach, we apply the proposed method to three representative models and conduct experiments on the WMT21 benchmarks.The results obtained from both in-domain and out-of-distribution evaluations consistently demonstrate improvements in correlation performance across different models.Furthermore, we provide evidence that the proposed method effectively reduces model uncertainty.The code is publicly available at https://github.com/NLP2CT/TaU.

Visual spatial description (VSD) aims to generate texts that describe the spatial relations of the given objects within images.Existing VSD work merely models the 2D geometrical vision features, thus inevitably falling prey to the problem of skewed spatial understanding of target objects.In this work, we investigate the incorporation of 3D scene features for VSD.With an external 3D scene extractor, we obtain the 3D objects and scene features for input images, based on which we construct a target object-centered 3D spatial scene graph (GO3D-S 2 G), such that we model the spatial semantics of target objects within the holistic 3D scenes.Besides, we propose a scene subgraph selecting mechanism, sampling topologically-diverse subgraphs from GO3D-S 2 G, where the diverse local structure features are navigated to yield spatially-diversified text generation.Experimental results on two VSD datasets demonstrate that our framework outperforms the baselines significantly, especially improving on the cases with complex visual spatial relations.Meanwhile, our method can produce more spatially-diversified generation.Code is available at https://github.com/zhaoyucs/VSD.

Token dropping is a recently-proposed strategy to speed up the pretraining of masked language models, such as BERT, by skipping the computation of a subset of the input tokens at several middle layers.It can effectively reduce the training time without degrading much performance on downstream tasks.However, we empirically find that token dropping is prone to a semantic loss problem and falls short in handling semantic-intense tasks ( §2).Motivated by this, we propose a simple yet effective semantic-consistent learning method (SCTD) to improve the token dropping.SCTD aims to encourage the model to learn how to preserve the semantic information in the representation space.Extensive experiments on 12 tasks show that, with the help of our SCTD, token dropping can achieve consistent and significant performance gains across all task types and model sizes.More encouragingly, SCTD saves up to 57% of pretraining time and brings up to +1.56% average improvement over the vanilla token dropping.

Tong Zhu, Junfei Ren, Zijian Yu, Mengsong Wu, Guoliang Zhang, Xiaoye Qu, Wenliang Chen, Zhefeng Wang, Baoxing Huai, Min Zhang. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

Entity linking aims to link ambiguous mentions to their corresponding entities in a knowledge base.One of the key challenges comes from insufficient labeled data for specific domains.Although dense retrievers have achieved excellent performance on several benchmarks, their performance decreases significantly when only a limited amount of in-domain labeled data is available.In such few-shot setting, we revisit the sparse retrieval method, and propose an ELECTRA-based keyword extractor to denoise the mention context and construct a better query expression.For training the extractor, we propose a distant supervision method to automatically generate training data based on overlapping tokens between mention contexts and entity descriptions.Experimental results on the ZESHEL dataset demonstrate that the proposed method outperforms state-of-the-art models by a significant margin across all test domains, showing the effectiveness of keywordenhanced sparse retrieval.Code is available at https://github.com/HITsz-TMG/

Self-training has proven to be an effective approach for cross-domain tasks, and in this study, we explore its application to cross-domain constituency parsing.Traditional self-training methods rely on limited and potentially lowquality raw corpora.To overcome this limitation, we propose enhancing self-training with the large language model (LLM) to generate domain-specific raw corpora iteratively.For the constituency parsing, we introduce grammar rules that guide the LLM in generating raw corpora and establish criteria for selecting pseudo instances.Our experimental results demonstrate that self-training for constituency parsing, equipped with an LLM, outperforms traditional methods regardless of the LLM's performance.Moreover, the combination of grammar rules and confidence criteria for pseudo-data selection yields the highest performance in the crossdomain constituency parsing 1 .

In multilingual translation research, the comprehension and utilization of language families are of paramount importance.Nevertheless, clustering languages based solely on their ancestral families can yield suboptimal results due to variations in the datasets employed during the model's training phase.To mitigate this challenge, we introduce an innovative method that leverages the fisher information matrix (FIM) to cluster language families, anchored on the multilingual translation model's characteristics.We hypothesize that language pairs with similar effects on model parameters exhibit a considerable degree of linguistic congruence and should thus be grouped cohesively.This concept has led us to define pseudo language families.We provide an in-depth discussion regarding the inception and application of these pseudo language families.Empirical evaluations reveal that employing these pseudo language families enhances performance over conventional language families in adapting a multilingual translation model to unfamiliar language pairs.The proposed methodology may also be extended to scenarios requiring language similarity measurements.The source code and associated scripts can be accessed at https://github.com/ecoli-hit/PseudoFamily.

Simultaneous machine translation (SiMT) models are trained to strike a balance between latency and translation quality.However, training these models to achieve high quality while maintaining low latency often leads to a tendency for aggressive anticipation.We argue that such issue stems from the autoregressive architecture upon which most existing SiMT models are built.To address those issues, we propose non-autoregressive streaming Transformer (NAST) which comprises a unidirectional encoder and a non-autoregressive decoder with intra-chunk parallelism.We enable NAST to generate the blank token or repetitive tokens to adjust its READ/WRITE strategy flexibly, and train it to maximize the nonmonotonic latent alignment with an alignmentbased latency loss.Experiments on various SiMT benchmarks demonstrate that NAST outperforms previous strong autoregressive SiMT baselines.

Conventional knowledge distillation (KD) approaches are commonly employed to compress neural machine translation (NMT) models.However, they only obtain one lightweight student each time.Consequently, we have to conduct KD multiple times when different students are required at the same time, which could be resource-intensive.Additionally, these students are individually optimized, and thus lack interactions with each other, leading to their potential not being fully exerted.In this work, we propose a novel All-In-One Knowledge Distillation (AIO-KD) framework for NMT, which generates multiple satisfactory students at once.Under AIO-KD, we first randomly extract fewer-layer subnetworks from the teacher as the sample students.Then, we jointly optimize the teacher and these students, where the students simultaneously learn the knowledge from the teacher and interact with other students via mutual learning.When utilized, we re-extract the candidate students, satisfying the specifications of various devices.Particularly, we adopt carefully-designed strategies for AIO-KD: 1) we dynamically detach gradients to prevent poorly-performed students from negatively affecting the teacher during the knowledge transfer, which could subsequently impact other students; 2) we design a twostage mutual learning strategy, which alleviates the negative impacts of poorly-performed students on the early-stage student interactions.Extensive experiments and in-depth analyses on three benchmarks demonstrate the effectiveness and eco-friendliness of AIO-KD.Our source code is available at https://github. com/DeepLearnXMU/AIO-KD.

Although pre-trained language models (PLM) have achieved great success in question answering (QA), their robustness is still insufficient to support their practical applications, especially in the face of distribution shifts.Recently, testtime adaptation (TTA) has shown great potential for solving this problem, which adapts the model to fit the test samples at test time.However, TTA sometimes causes model collapse, making almost all the model outputs incorrect, which has raised concerns about its stability and reliability.In this paper, we delve into why TTA causes model collapse and find that the imbalanced label distribution inherent in QA is the reason for it.To address this problem, we propose Anti-Collapse Fast test-time adaptation (Anti-CF), which utilizes the source model's output to regularize the update of the adapted model during test time.We further design an efficient side block to reduce its inference time.Extensive experiments on various distribution shift scenarios and pre-trained language models (e.g., XLM-RoBERTa, BLOOM) demonstrate that our method can achieve comparable or better results than previous TTA methods at a speed close to vanilla forward propagation, which is 1.8× to 4.4× speedup compared to previous TTA methods.Our code is available at https://github.com/yisunlp/Anti-CF.

Beyond the great cognitive powers showcased by language models, it is crucial to scrutinize whether their reasoning capabilities stem from strong generalization or merely exposure to relevant data.As opposed to constructing increasingly complex logic, this paper probes into the boolean logic, the root capability of a logical reasoner.We find that any pre-trained language models even including large language models only behave like a random selector in the face of multi-nested boolean logic, a task that humans can handle with ease.To empower language models with this fundamental capability, this paper proposes a new self-supervised learning method Curriculum Logical Reasoning (CLR), where we augment the training data with nested boolean logic chain step-by-step, and program the training from simpler logical patterns gradually to harder ones.This new training paradigm allows language models to effectively generalize to much harder and longer-hop logic, which can hardly be learned through naive training.Furthermore, we show that boolean logic is a great foundation for improving the subsequent general logical tasks 1 .

An end-to-end speech-to-text (S2T) translation model is usually initialized from a pretrained speech recognition encoder and a pretrained text-to-text (T2T) translation decoder.Although this straightforward setting has been shown empirically successful, there do not exist clear answers to the research questions: 1) how are speech and text modalities fused in S2T model and 2) how to better fuse the two modalities?In this paper, we take the first step toward understanding the fusion of speech and text features in S2T model.We first design and release a 10GB linguistic probing benchmark, namely Speech-Senteval, to investigate the acoustic and linguistic behaviors of S2T models.Preliminary analysis reveals that the uppermost encoder layers of the S2T model can not learn linguistic knowledge efficiently, which is crucial for accurate translation.Based on the finding, we further propose a simple plug-in prompt-learning strategy on the uppermost encoder layers to broaden the abstract representation power of the encoder of S2T models.We call such a promptenhanced S2T model PromptST.Experimental results on four widely-used S2T datasets show that PromptST can deliver significant improvements over a strong baseline by capturing richer linguistic knowledge.Benchmarks,

Large language models (LLMs) have demonstrated exceptional performance in reasoning tasks with dedicated Chain-of-Thought (CoT) prompts.Further enhancing CoT prompts with exquisite exemplars can significantly improve reasoning performance.However, the effectiveness of CoT prompts may fluctuate dramatically with different choices of in-context examples.Additionally, manual construction of rationale steps can be time-consuming, presenting challenges for the widespread adoption of CoT prompting.In this work, we propose a novel approach by introducing information entropy (IE) as a criteria on for CoT prompt selection.We extend this criterion to the CoT generation and inference stages, automatically generating CoT prompts with higher information entropy scores and adaptively determining the number of samples.These three stages together form our proposed information entropy based multi-step reasoning for large language models, named INFORM.Our experiments across seven reasoning benchmarks utilizing two language models(GPT-3.5-Turboand text-davinci-003) demonstrate the superiority of INFORM both in performance and efficiency. 1

Existing approaches to Emotion Recognition in Conversation (ERC) use a fixed context window to recognize speakers’ emotion, which may lead to either scantiness of key context or interference of redundant context. In response, we explore the benefits of variable-length context and propose a more effective approach to ERC. In our approach, we leverage different context windows when predicting the emotion of different utterances. New modules are included to realize variable-length context: 1) two speaker-aware units, which explicitly model inner- and inter-speaker dependencies to form distilled conversational context and 2) a top-k normalization layer, which determines the most proper context windows from the conversational context to predict emotion. Experiments and ablation study show that our approach outperforms several strong baselines on three public datasets.

Existing skeleton-based action recognition methods typically follow a centralized learning paradigm, which can pose privacy concerns when exposing human-related videos. Federated Learning (FL) has attracted much attention due to its outstanding advantages in privacy-preserving. However, directly applying FL approaches to skeleton videos suffers from unstable training. In this paper, we investigate and discover that the heterogeneous human topology graph structure is the crucial factor hindering training stability. To address this limitation, we pioneer a novel Federated Skeleton-based Action Recognition (FSAR) paradigm, which enables the construction of a globally generalized model without accessing local sensitive data. Specifically, we introduce an Adaptive Topology Structure (ATS), separating generalization and personalization by learning a domain-invariant topology shared across clients and a domain-specific topology decoupled from global model aggregation. Furthermore, we explore Multi-grain Knowledge Distillation (MKD) to mitigate the discrepancy between clients and server caused by distinct updating patterns through aligning shallow block-wise motion features. Extensive experiments on multiple datasets demonstrate that FSAR outperforms state-of-the-art FL-based methods while inherently protecting privacy.

Text-based person search aims to retrieve the specified person images given a textual description. The key to tackling such a challenging task is to learn powerful multi-modal representations. Towards this, we propose a Relation and Sensitivity aware representation learning method (RaSa), including two novel tasks: Relation-Aware learning (RA) and Sensitivity-Aware learning (SA). For one thing, existing methods cluster representations of all positive pairs without distinction and overlook the noise problem caused by the weak positive pairs where the text and the paired image have noise correspondences, thus leading to overfitting learning. RA offsets the overfitting risk by introducing a novel positive relation detection task (i.e., learning to distinguish strong and weak positive pairs). For another thing, learning invariant representation under data augmentation (i.e., being insensitive to some transformations) is a general practice for improving representation's robustness in existing methods. Beyond that, we encourage the representation to perceive the sensitive transformation by SA (i.e., learning to detect the replaced words), thus promoting the representation's robustness. Experiments demonstrate that RaSa outperforms existing state-of-the-art methods by 6.94%, 4.45% and 15.35% in terms of Rank@1 on CUHK-PEDES, ICFG-PEDES and RSTPReid datasets, respectively. Code is available at: https://github.com/Flame-Chasers/RaSa.

Event Causality Identification (ECI) aims to identify the causality between a pair of event mentions in a document, which is composed of sentence-level ECI (SECI) and document-level ECI (DECI). Previous work applies various reasoning models to identify the implicit event causality. However, they indiscriminately reason all event causality in the same way, ignoring that most inter-sentence event causality depends on intra-sentence event causality to infer. In this paper, we propose a progressive graph pairwise attention network (PPAT) to consider the above dependence. PPAT applies a progressive reasoning strategy, as it first predicts the intra-sentence event causality, and then infers the more implicit inter-sentence event causality based on the SECI result. We construct a sentence boundary event relational graph, and PPAT leverages a simple pairwise attention mechanism, which attends to different reasoning chains on the graph. In addition, we propose a causality-guided training strategy for assisting PPAT in learning causality-related representations on every layer. Extensive experiments show that our model achieves state-of-the-art performance on three benchmark datasets (5.5%, 2.2% and 4.5% F1 gains on EventStoryLine, MAVEN-ERE and Causal-TimeBank). Code is available at https://github.com/HITsz-TMG/PPAT.

Sequential recommendation methods are increasingly important in cutting-edge recommender systems. Through leveraging historical records, the systems can capture user interests and perform recommendations accordingly. State-of-the-art sequential recommendation models proposed very recently combine contrastive learning techniques for obtaining high-quality user representations. Though effective and performing well, the models based on contrastive learning require careful selection of data augmentation methods and pretext tasks, efficient negative sampling strategies, and massive hyper-parameters validation. In this paper, we propose an ultra-simple alternative for obtaining better user representations and improving sequential recommendation performance. Specifically, we present a simple yet effective Consistency T braining method for sequential Recommendation (CT4Rec) in which only two extra training objectives are utilized without any structural modifications and data augmentation. Experiments on three benchmark datasets and one large newly crawled industrial corpus demonstrate that our proposed method outperforms SOTA models by a large margin and with much less training time than these based on contrastive learning. Online evaluation on real-world content recommendation system also achieves 2.717% improvement on the click-through rate and 3.679% increase on the average click number per capita. Further exploration reveals that such a simple method has great potential for CTR prediction. Our code is available at https://github.com/ct4rec/CT4Rec.git.

Text-based person search (TBPS) aims to retrieve the images of the target person from a large image gallery based on a given natural language description. Existing methods are dominated by training models with parallel image-text pairs, which are very costly to collect. In this paper, we make the first attempt to explore TBPS without parallel image-text data (μ-TBPS), in which only non-parallel images and texts, or even image-only data, can be adopted. Towards this end, we propose a two-stage framework, generation-then-retrieval (GTR), to first generate the corresponding pseudo text for each image and then perform the retrieval in a supervised manner. In the generation stage, we propose a fine-grained image captioning strategy to obtain an enriched description of the person image, which firstly utilizes a set of instruction prompts to activate the off-the-shelf pretrained vision-language model to capture and generate fine-grained person attributes, and then converts the extracted attributes into a textual description via the finetuned large language model or the hand-crafted template. In the retrieval stage, considering the noise interference of the generated texts for training model, we develop a confidence score-based training scheme by enabling more reliable texts to contribute more during the training. Experimental results on multiple TBPS benchmarks (i.e., CUHK-PEDES, ICFG-PEDES and RSTPReid) show that the proposed GTR can achieve a promising performance without relying on parallel image-text data.

Masked video modeling has shown remarkable performance in downstream tasks by predicting masked video tokens from visible ones. However, training models from scratch on large-scale unlabeled data remains computationally challenging and time-consuming. Moreover, the commonly used random-based sampling techniques may lead to the selection of redundant or low-information regions, hindering the model from learning discriminative representations within the limited training epochs. To achieve efficient pre-training, we propose MaskAgain, an efficient feature-based knowledge distillation framework for masked video pre-training that facilitates knowledge transfer from a pre-trained teacher model to a student model. In contrast to previous approaches that align all visible token features with the teacher model at output layers, MaskAgain adopts a selective approach by masking visible tokens again at both the hidden and output layers of the transformer block. Attention mechanisms are utilized for informative feature selection. At the hidden level, attention maps generated by the transformer's multi-head attention structure are utilized to select crucial token information at both temporally-global and temporally-local levels. Additionally, at the output level, an activation-based attention map is generated using token features, enabling us to focus on important tokens while preserving feature similarity and the relationship matrix similarity between patches. Extensive experimental results show that MaskAgain achieves comparable or even better performance than existing methods on benchmark datasets with much fewer training epochs and much less memory, which demonstrates that MaskAgain allows for efficient pre-training of accurate video models, reducing computational resources and training time significantly. Code is released at https://github.com/xiaojieli0903/MaskAgain.

Self-supervised learning methods have shown significant promise in acquiring robust spatiotemporal representations from unlabeled videos. In this work, we address three critical limitations in existing self-supervised video representation learning: 1) insufficient utilization of contextual information and lifelong memory, 2) lack of fine-grained visual concept alignment, and 3) neglect of the feature distribution gap between encoders. To overcome these limitations, we propose a novel memory-enhanced predictor that leverages key-value memory networks with separate memories for the online and target encoders. This design enables the effective storage and retrieval of contextual knowledge, facilitating informed predictions and enhancing overall performance. Additionally, we introduce a visual concept alignment module that ensures fine-grained alignment of shared semantic information across segments of the same video. By employing coupled dictionary learning, we effectively decouple visual concepts, enriching the semantic representation stored in the memory networks. Our proposed approach is extensively evaluated on widely recognized benchmarks for action recognition and retrieval tasks, demonstrating its superiority in learning generalized video representations with significantly improved performance compared to existing state-of-the-art self-supervised learning methods. Code is released at https://github.com/xiaojieli0903/FGKVMemPred_video.

Multi-modal multi-hop question answering involves answering a question by reasoning over multiple input sources from different modalities. Existing methods often retrieve evidences separately and then use a language model to generate an answer based on the retrieved evidences, and thus do not adequately connect candidates and are unable to model the interdependent relations during retrieval. Moreover, the pipelined approaches of retrieval and generation might result in poor generation performance when retrieval performance is low. To address these issues, we propose a Structured Knowledge and Unified Retrieval-Generation (SKURG) approach. SKURG employs an Entity-centered Fusion Encoder to align sources from different modalities using shared entities. It then uses a unified Retrieval-Generation Decoder to integrate intermediate retrieval results for answer generation and also adaptively determine the number of retrieval steps. Extensive experiments on two representative multi-modal multi-hop QA datasets MultimodalQA and WebQA demonstrate that SKURG outperforms the state-of-the-art models in both source retrieval and answer generation performance with fewer parameters1.

To achieve multi-modal intelligence, AI must be able to process and respond to inputs from multimodal sources. However, many current question answering models are limited to specific types of answers, such as yes/no and number, and require additional human assessments. Recently, Visual-Text Question Answering (VQTA) dataset has been proposed to fix this gap. In this paper, we conduct an exhaustive analysis and exploration of this task. Specifically, we implement a T5-based multi-modal generative network that overcomes the limitations of traditional labeling space and provides more freedom in responses. Our approach achieve the best performance in both English and Chinese tracks in the VTQA challenge.

As one of the core video semantic understanding tasks, Video Semantic Role Labeling (VidSRL) aims to detect the salient events from given videos, by recognizing the predict-argument event structures and the interrelationships between events. While recent endeavors have put forth methods for VidSRL, they can be mostly subject to two key drawbacks, including the lack of fine-grained spatial scene perception and the insufficiently modeling of video temporality. Towards this end, this work explores a novel holistic spatio-temporal scene graph (namely HostSG) representation based on the existing dynamic scene graph structures, which well model both the fine-grained spatial semantics and temporal dynamics of videos for VidSRL. Built upon the HostSG, we present a nichetargeting VidSRL framework. A scene-event mapping mechanism is first designed to bridge the gap between the underlying scene structure and the high-level event semantic structure, resulting in an overall hierarchical scene-event (termed ICE) graph structure. We further perform iterative structure refinement to optimize the ICE graph, e.g., filtering noisy branches and newly building informative connections, such that the overall structure representation can best coincide with end task demand. Finally, three subtask predictions of VidSRL are jointly decoded, where the end-to-end paradigm effectively avoids error propagation. On the benchmark dataset, our framework boosts significantly over the current best-performing model. Further analyses are shown for a better understanding of the advances of our methods. Our HostSG representation shows greater potential to facilitate a broader range of other video understanding tasks.

Video-based visible-infrared person re-identification (VVI-ReID) aims to retrieve video sequences of the same pedestrian from different modalities. The key of VVI-ReID is to learn discriminative sequence-level representations that are invariant to both intra- and inter-modal discrepancies. However, most works only focus on the elimination of modality-gap while ignore the distractors within the modality. Moreover, existing sequence-level representation learning approaches are limited to a single video, failing to mine the correlations among multiple videos of the same pedestrian. In this paper, we propose a Style Augmentation, Attack and Defense network with Graph-based dual interaction (SAADG) to guarantee the semantic consistency against both intra-modal discrepancies and inter-modal gap. Specifically, we first generate diverse styles for video frames by random style variation in image spaces. Followed by the style attack and defense, the intra- and inter-modal discrepancies are modeled as different types of style disturbance (attack), and our model achieves to keep the id-related content invariant under such attack. Besides, a graph-based dual interaction module is further introduced to fully explore the cross-view and cross-modal correlations among various videos of the same identity, which are then transferred to the sequence-level representations. Extensive experiments on the public SYSU-MM01 and HITSZ-VCM datasets show that our approach achieves the remarkable performance compared with state-of-the-arts. The code is available at https://github.com/ChuhaoZhou99/SAADG_VVIReID.

Maximum likelihood estimation (MLE) is a statistical method used to estimate the parameters of a probability distribution that best explain the observed data. In the context of text generation, MLE is often used to train generative language models, which can then be used to generate new text. However, we argue that MLE is not always necessary and optimal, especially for closed-ended text generation tasks like machine translation. In these tasks, the goal of model is to generate the most appropriate response, which does not necessarily require it to estimate the entire data distribution with MLE. To this end, we propose a novel class of training objectives based on convex functions, which enables text generation models to focus on highly probable outputs without having to estimate the entire data distribution. We investigate the theoretical properties of the optimal predicted distribution when applying convex functions to the loss, demonstrating that convex functions can sharpen the optimal distribution, thereby enabling the model to better capture outputs with high probabilities. Experiments on various text generation tasks and models show the effectiveness of our approach. It enables autoregressive models to bridge the gap between greedy and beam search, and facilitates the learning of non-autoregressive models with a maximum improvement of 9+ BLEU points. Moreover, our approach also exhibits significant impact on large language models (LLMs), substantially enhancing their generative capability on various tasks. Source code is available at \url{https://github.com/ictnlp/Convex-Learning}.

Non-autoregressive (NAR) generation, which is first proposed in neural machine translation (NMT) to speed up inference, has attracted much attention in both machine learning and natural language processing communities. While NAR generation can significantly accelerate inference speed for machine translation, the speedup comes at the cost of sacrificed translation accuracy compared to its counterpart, autoregressive (AR) generation. In recent years, many new models and algorithms have been designed/proposed to bridge the accuracy gap between NAR generation and AR generation. In this paper, we conduct a systematic survey with comparisons and discussions of various non-autoregressive translation (NAT) models from different aspects. Specifically, we categorize the efforts of NAT into several groups, including data manipulation, modeling methods, training criterion, decoding algorithms, and the benefit from pre-trained models. Furthermore, we briefly review other applications of NAR models beyond machine translation, such as grammatical error correction, text summarization, text style transfer, dialogue, semantic parsing, automatic speech recognition, and so on. In addition, we also discuss potential directions for future exploration, including releasing the dependency of KD, reasonable training objectives, pre-training for NAR, and wider applications, etc. We hope this survey can help researchers capture the latest progress in NAR generation, inspire the design of advanced NAR models and algorithms, and enable industry practitioners to choose appropriate solutions for their applications.

Neural chat translation (NCT) aims to translate a cross-lingual chat between speakers of different languages. Existing context-aware NMT models cannot achieve satisfactory performances due to the following inherent problems: 1) limited resources of annotated bilingual dialogues; 2) the neglect of modelling conversational properties; 3) training discrepancy between different stages. To address these issues, in this paper, we propose a multi-task multi-stage transitional (MMT) training framework, where an NCT model is trained using the bilingual chat translation dataset and additional monolingual dialogues. We elaborately design two auxiliary tasks, namely utterance discrimination and speaker discrimination, to introduce the modelling of dialogue coherence and speaker characteristic into the NCT model. The training process consists of three stages: 1) sentence-level pre-training on large-scale parallel corpus; 2) intermediate training with auxiliary tasks using additional monolingual dialogues; 3) context-aware fine-tuning with gradual transition. Particularly, the second stage serves as an intermediate phase that alleviates the training discrepancy between the pre-training and fine-tuning stages. Moreover, to make the stage transition smoother, we train the NCT model using a gradual transition strategy, i.e., gradually transiting from using monolingual to bilingual dialogues. Extensive experiments on two language pairs demonstrate the effectiveness and superiority of our proposed training framework.

Syntactic processing is fundamental to natural language processing. It provides rich and comprehensive syntax information in sentences that could be potentially beneficial for downstream tasks. Recently, pretrained language models have shown great success in Chinese syntactic processing, which typically involves word segmentation, POS tagging, and dependency parsing. However, the on-going research never ends since performance would be degraded drastically when tested on a highly-discrepant domain. This problem is widely accepted as domain adaptation, where the test domain differs from the training domain in supervised learning. Self-training is one promising solution for it, and straightforward source-to-target adaptation has already shown remarkable effectiveness in previous work. While this strategy ignores the fact that sentences of the target domain sentences may have very different gaps from the source training domain. More specifically, sentences with large gaps might fail by direct self-training adaptation. To this end, we propose fine-grained domain adaptation for Chinese syntactic processing in this work, aiming to model the gaps between the source and the target domains accurately and progressively. The key idea is to divide the target domain into fine-grained subdomains by using a specified domain distance metric, and then perform gradual self-training on the subdomains. We further offer an intuitive theoretical illustration based on the theory of Kumar et al. (2020) approximately. In addition, a novel representation learning framework is proposed to encode fine-grained subdomains effectively, aiming to utilize the above idea fully. Experimental results on benchmark datasets show that our method can achieve significant improvements over a variety of baselines.

Unsupervised cross-lingual transfer has been shown great potentials for dependency parsing of the low-resource languages when there is no annotated treebank available. Recently, the self-training method has received increasing interests because of its state-of-the-art performance in this scenario. In this work, we advance the method further by coupling it with curriculum learning, which guides the self-training in an easy-to-hard manner. Concretely, we present a novel metric to measure the instance difficulty of a dependency parser which is trained mainly on a Treebank from a resource-rich source language. By using the metric, we divide a low-resource target language into several fine-grained sub-languages by their difficulties, and then apply iterative-self-training progressively on these sub-languages. To fully explore the auto-parsed training corpus from sub-languages, we exploit an improved parameter generation network to model the sub-languages for better representation learning. Experimental results show that our final curriculum-style self-training can outperform a range of strong baselines, leading to new state-of-the-art results on unsupervised cross-lingual dependency parsing. We also conduct detailed experimental analyses to examine the proposed approach in depth for comprehensive understandings.

Directly training a document-to-document (Doc2Doc) neural machine translation (NMT) via Transformer from scratch, especially on small datasets, usually fails to converge. Our dedicated probing tasks show that 1) both the absolute position and relative position information gets gradually weakened or even vanished once it reaches the upper encoder layers, and 2) the vanishing of absolute position information in encoder output causes the training failure of Doc2Doc NMT. To alleviate this problem, we propose a position-aware Transformer (P-Transformer) to enhance both the absolute and relative position information in both self-attention and cross-attention. Specifically, we integrate absolute positional information, i.e., position embeddings, into the query-key pairs both in self-attention and cross-attention through a simple yet effective addition operation. Moreover, we also integrate relative position encoding in self-attention. The proposed P-Transformer utilizes sinusoidal position encoding and does not require any task-specified position embedding, segment embedding, or attention mechanism. Through the above methods, we build a Doc2Doc NMT model with P-Transformer, which ingests the source document and completely generates the target document in a sequence-to-sequence (seq2seq) way. In addition, P-Transformer can be applied to seq2seq-based document-to-sentence (Doc2Sent) and sentence-to-sentence (Sent2Sent) translations. Extensive experimental results of Doc2Doc NMT show that P-Transformer significantly outperforms strong baselines on the widely-used 9 document-level datasets in 7 language pairs, covering small-, middle-, and large-scales, and achieves a new state-of-the-art. Experimentation on discourse phenomena shows that our Doc2Doc NMT models improve the translation quality in both BLEU and discourse coherence. We make our code available on Github.

Neural machine translation uses a decoder to generate target words auto-regressively by predicting the next target word conditioned on a given source sentence and its previously predicted target words, i.e, its translation history, which suffers from two limitations: 1) the prediction of next word depends heavily on the quality of its history information. Moreover, the discrepancy between training and inference exacerbates this limitation; 2) this left-to-right decoding way cannot make full use of the target-side future information, which leads to the issue of unbalanced outputs. On the one hand, we alleviate the first limitation with a history-refining module, which learns to examine the quality of each history word by assigning it a confidence score. The confidence score is further used as a gate to control the amount of its word embedding flowing to the decoder. On the other hand, we attack the second limitation with a future-foreseeing module, which learns the distribution of future translation at each decoding time step. More importantly, we further propose refining history for future-aware NMT since the two modules can be closely incorporated as they focus on different kinds of context. Experimental results on various translation tasks with different scaled datasets, including WMT English$\leftrightarrow${German, French, Romanian}, show that our proposed approach achieves significant improvements over strong Transformer-based NMT baselines.

Pre-training and fine-tuning have become the de facto paradigm in many natural language processing (NLP) tasks. However, compared to other NLP tasks, neural machine translation (NMT) aims to generate target language sentences through the contextual representation from the source language counterparts. This characteristic means the optimization objective of NMT is far from that of the universal pre-trained models (PTMs), leading to the standard procedure of pre-training and fine-tuning does not work well in NMT. In this paper, we propose a novel framework to deep fuse the pre-trained representation into NMT, fully exploring the potential of PTMs in NMT. Specifically, we directly replace the randomly initialized Transformer encoder with a pre-trained encoder and propose a layer-wise coordination structure to coordinate PTM and NMT decoder learning. Then, we introduce a partitioned multi-task learning method to fine-tune the pre-trained parameter, reducing the gap between PTM and NMT by progressively learning the task-specific representation. Experimental results show that our approach achieves considerable improvements on WMT14 En2De, WMT14 En2Fr, and WMT16 Ro2En translation benchmarks and outperforms previous work in both autoregressive and non-autoregressive NMT models.

Implicit discourse relation recognition (IDRR) is a challenging but crucial task in discourse analysis. Most existing methods train multiple models to predict multi-level labels independently, while ignoring the dependence between hierarchically structured labels. In this paper, we consider multi-level IDRR as a conditional label sequence generation task and propose a Label Dependence-aware Sequence Generation Model (LDSGM) for it. Specifically, we first design a label attentive encoder to learn the global representation of an input instance and its level-specific contexts, where the label dependence is integrated to obtain better label embeddings. Then, we employ a label sequence decoder to output the predicted labels in a top-down manner, where the predicted higher-level labels are directly used to guide the label prediction at the current level. We further develop a mutual learning enhanced training method to exploit the label dependence in a bottom-up direction, which is captured by an auxiliary decoder introduced during training. Experimental results on the PDTB dataset show that our model achieves the state-of-the-art performance on multi-level IDRR. We release our code at https://github.com/nlpersECJTU/LDSGM.

Regularization methods applying input perturbation have drawn considerable attention and have been frequently explored for NMT tasks in recent years.Despite their simplicity and effectiveness, we argue that these methods are limited by the under-fitting of training data.In this paper, we utilize prediction difference for ground-truth tokens to analyze the fitting of token-level samples and find that underfitting is almost as common as over-fitting.We introduce prediction difference regularization (PD-R), a simple and effective method that can reduce over-fitting and under-fitting at the same time.For all token-level samples, PD-R minimizes the prediction difference between the original pass and the input-perturbed pass, making the model less sensitive to small input changes, thus more robust to both perturbations and under-fitted training data.Experiments on three widely used WMT translation tasks show that our approach can significantly improve over existing perturbation regularization methods.On WMT16 En-De task, our model achieves 1.80 SacreBLEU improvement over vanilla transformer.

Bei Li, Quan Du, Tao Zhou, Yi Jing, Shuhan Zhou, Xin Zeng, Tong Xiao, JingBo Zhu, Xuebo Liu, Min Zhang. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Supervised parsing models have achieved impressive results on in-domain texts.However, their performances drop drastically on out-ofdomain texts due to the data distribution shift.The shared-private model has shown its promising advantages for alleviating this problem via feature separation, whereas prior works pay more attention to enhancing shared features but neglect the in-depth relevance of specific ones.To address this issue, we for the first time apply a dynamic matching network on the shared-private model for semi-supervised crossdomain dependency parsing.Meanwhile, considering the scarcity of target-domain labeled data, we leverage unlabeled data from two aspects, i.e., designing a new training strategy to improve the capability of the dynamic matching network and fine-tuning BERT to obtain domain-related contextualized representations.Experiments on benchmark datasets show that our proposed model consistently outperforms various baselines, leading to new state-of-theart results on all domains.Detailed analysis on different matching strategies demonstrates that it is essential to learn suitable matching weights to emphasize useful features and ignore useless or even harmful ones.Besides, our proposed model can be directly extended to multi-source domain adaptation and achieves best performances among various baselines, further verifying the effectiveness and robustness.

Pre-trained language models (PLMs) have shown great potentials in natural language processing (NLP) including rhetorical structure theory (RST) discourse parsing.Current PLMs are obtained by sentence-level pre-training, which is different from the basic processing unit, i.e. element discourse unit (EDU).To this end, we propose a second-stage EDU-level pretraining approach in this work, which presents two novel tasks to learn effective EDU representations continually based on well pre-trained language models.Concretely, the two tasks are (1) next EDU prediction (NEP) and ( 2) discourse marker prediction (DMP).We take a state-of-the-art transition-based neural parser as baseline, and adopt it with a light bi-gram EDU modification to effectively explore the EDU-level pre-trained EDU representation.Experimental results on a benckmark dataset show that our method is highly effective, leading a 2.1-point improvement in F1-score.All codes and pre-trained models will be released publicly to facilitate future studies. 1

Recent works of opinion expression identification (OEI) rely heavily on the quality and scale of the manually-constructed training corpus, which could be extremely difficult to satisfy.Crowdsourcing is one practical solution for this problem, aiming to create a large-scale but quality-unguaranteed corpus.In this work, we investigate Chinese OEI with extremelynoisy crowdsourcing annotations, constructing a dataset at a very low cost.Following Zhang et al. (2021), we train the annotator-adapter model by regarding all annotations as goldstandard in terms of crowd annotators, and test the model by using a synthetic expert, which is a mixture of all annotators.As this annotatormixture for testing is never modeled explicitly in the training phase, we propose to generate synthetic training samples by a pertinent mixup strategy to make the training and testing highly consistent.The simulation experiments on our constructed dataset show that crowdsourcing is highly promising for OEI, and our proposed annotator-mixup can further enhance the crowdsourcing modeling.

Most dominant neural machine translation (NMT) models are restricted to make predictions only according to the local context of preceding words in a left-to-right manner.Although many previous studies try to incorporate global information into NMT models, there still exist limitations on how to effectively exploit bidirectional global context.In this paper, we propose a Confidence Based Bidirectional Global Context Aware (CB-BGCA) training framework for NMT, where the NMT model is jointly trained with an auxiliary conditional masked language model (CMLM).The training consists of two stages:(1) multi-task joint training; (2) confidence based knowledge distillation.At the first stage, by sharing encoder parameters, the NMT model is additionally supervised by the signal from the CMLM decoder that contains bidirectional global contexts.Moreover, at the second stage, using the CMLM as teacher, we further pertinently incorporate bidirectional global context to the NMT model on its unconfidently-predicted target words via knowledge distillation.Experimental results show that our proposed CB-BGCA training framework significantly improves the NMT model by +1.02, +1.30 and +0.57BLEU scores on three large-scale translation datasets, namely WMT'14 Englishto-German, WMT'19 Chinese-to-English and WMT'14 English-to-French, respectively.

The existing research efforts in Multimodal Sentiment Analysis (MSA) have focused on developing the expressive ability of neural networks to fuse information from different modalities. However, these approaches lack a mechanism to understand the complex relations within and across different modalities, since some sentiments may be scattered in different modalities. To this end, in this paper, we propose a novel hierarchical graph contrastive learning (HGraph-CL) framework for MSA, aiming to explore the intricate relations of intra- and inter-modal representations for sentiment extraction. Specifically, regarding the intra-modal level, we build a unimodal graph for each modality representation to account for the modality-specific sentiment implications. Based on it, a graph contrastive learning strategy is adopted to explore the potential relations based on unimodal graph augmentations. Furthermore, we construct a multimodal graph of each instance based on the unimodal graphs to grasp the sentiment relations between different modalities. Then, in light of the multimodal augmentation graphs, a graph contrastive learning strategy over the inter-modal level is proposed to ulteriorly seek the possible graph structures for precisely learning sentiment relations. This essentially allows the framework to understand the appropriate graph structures for learning intricate relations among different modalities. Experimental results on two benchmark datasets show that the proposed framework outperforms the state-of-the-art baselines in MSA.

Most existing methods on robust neural machine translation (NMT) construct adversarial examples by injecting noise into authentic examples and indiscriminately exploit two types of examples. They require the model to translate both the authentic source sentence and its adversarial counterpart into the identical target sentence within the same training stage, which may be a suboptimal choice to achieve robust NMT. In this paper, we first conduct a preliminary study to confirm this claim and further propose an Iterative Scheduled Data-switch Training Framework to mitigate this problem. Specifically, we introduce two training stages, iteratively switching between authentic and adversarial examples. Compared with previous studies, our model focuses more on just one type of examples at each single stage, which can better exploit authentic and adversarial examples, and thus obtaining a better robust NMT model. Moreover, we introduce an improved curriculum learning method with a sampling strategy to better schedule the process of noise injection. Experimental results show that our model significantly surpasses several competitive baselines on four translation benchmarks. Our source code is available at https://github.com/DeepLearnXMU/RobustNMT-ISDST.

The conventional success of textual classification relies on annotated data, and the new paradigm of pre-trained language models (PLMs) still requires a few labeled data for downstream tasks. However, in real-world applications, label noise inevitably exists in training data, damaging the effectiveness, robustness, and generalization of the models constructed on such data. Recently, remarkable achievements have been made to mitigate this dilemma in visual data, while only a few explore textual data. To fill this gap, we present SelfMix, a simple yet effective method, to handle label noise in text classification tasks. SelfMix uses the Gaussian Mixture Model to separate samples and leverages semi-supervised learning. Unlike previous works requiring multiple models, our method utilizes the dropout mechanism on a single model to reduce the confirmation bias in self-training and introduces a textual level mixup training strategy. Experimental results on three text classification benchmarks with different types of text show that the performance of our proposed method outperforms these strong baselines designed for both textual and visual data under different noise ratios and noise types. Our anonymous code is available at https://github.com/noise-learning/SelfMix.

Discourse parsing on multi-party dialogues is an important but difficult task in dialogue systems and conversational analysis. It is believed that speaker interactions are helpful for this task. However, most previous research ignores speaker interactions between different speakers. To this end, we present a speaker-aware model for this task. Concretely, we propose a speaker-context interaction joint encoding (SCIJE) approach, using the interaction features between different speakers. In addition, we propose a second-stage pre-training task, same speaker prediction (SSP), enhancing the conversational context representations by predicting whether two utterances are from the same speaker. Experiments on two standard benchmark datasets show that the proposed model achieves the best-reported performance in the literature. We will release the codes of this paper to facilitate future research.

Semantic role labeling (SRL) is a fundamental yet challenging task in the NLP community. Recent works of SRL mainly fall into two lines: 1) BIO-based; 2) span-based. Despite ubiquity, they share some intrinsic drawbacks of not considering internal argument structures, potentially hindering the model’s expressiveness. The key challenge is arguments are flat structures, and there are no determined subtree realizations for words inside arguments. To remedy this, in this paper, we propose to regard flat argument spans as latent subtrees, accordingly reducing SRL to a tree parsing task. In particular, we equip our formulation with a novel span-constrained TreeCRF to make tree structures span-aware and further extend it to the second-order case. We conduct extensive experiments on CoNLL05 and CoNLL12 benchmarks. Results reveal that our methods perform favorably better than all previous syntax-agnostic works, achieving new state-of-the-art under both end-to-end and w/ gold predicates settings.

This paper proposes to cast end-to-end span-based SRL as a word-based graph parsing task. The major challenge is how to represent spans at the word level. Borrowing ideas from research on Chinese word segmentation and named entity recognition, we propose and compare four different schemata of graph representation, i.e., BES, BE, BIES, and BII, among which we find that the BES schema performs the best. We further gain interesting insights through detailed analysis. Moreover, we propose a simple constrained Viterbi procedure to ensure the legality of the output graph according to the constraints of the SRL structure. We conduct experiments on two widely used benchmark datasets, i.e., CoNLL05 and CoNLL12. Results show that our word-based graph parsing approach achieves consistently better performance than previous results, under all settings of end-to-end and predicate-given, without and with pre-trained language models (PLMs). More importantly, our model can parse 669/252 sentences per second, without and with PLMs respectively.

This work proposes a syntax-enhanced grammatical error correction (GEC) approach named SynGEC that effectively incorporates dependency syntactic information into the encoder part of GEC models. 1 The key challenge for this idea is that off-the-shelf parsers are unreliable when processing ungrammatical sentences.To confront this challenge, we propose to build a tailored GEC-oriented parser (GOPar) using parallel GEC training data as a pivot.First, we design an extended syntax representation scheme that allows us to represent both grammatical errors and syntax in a unified tree structure.Then, we obtain parse trees of the source incorrect sentences by projecting trees of the target correct sentences.Finally, we train GOPar with such projected trees.For GEC, we employ the graph convolution network to encode source-side syntactic information produced by GOPar, and fuse them with the outputs of the Transformer encoder.Experiments on mainstream English and Chinese GEC datasets show that our proposed SynGEC approach consistently and substantially outperforms strong baselines and achieves competitive performance.

Pretraining-based (PT-based) automatic evaluation metrics (e.g., BERTScore and BARTScore) have been widely used in several sentence generation tasks (e.g., machine translation and text summarization) due to their better correlation with human judgments over traditional overlapbased methods.Although PT-based methods have become the de facto standard for training grammatical error correction (GEC) systems, GEC evaluation still does not benefit from pretrained knowledge.This paper takes the first step towards understanding and improving GEC evaluation with pretraining.We first find that arbitrarily applying PT-based metrics to GEC evaluation brings unsatisfactory correlation results because of the excessive attention to inessential systems outputs (e.g., unchanged parts).To alleviate the limitation, we propose a novel GEC evaluation metric to achieve the best of both worlds, namely PT-M 2 , which only uses PT-based metrics to score those corrected parts.Experimental results on the CoNLL14 evaluation task show that PT-M 2 significantly outperforms existing methods, achieving a new state-of-the-art result of 0.949 Pearson correlation.Further analysis reveals that PT-M 2 is robust to evaluate competitive GEC systems.

Boundary information is critical for various Chinese language processing tasks, such as word segmentation, part-of-speech tagging, and named entity recognition.Previous studies usually resorted to the use of a high-quality external lexicon, where lexicon items can offer explicit boundary information.However, to ensure the quality of the lexicon, great human effort is always necessary, which has been generally ignored.In this work, we suggest unsupervised statistical boundary information instead, and propose an architecture to encode the information directly into pre-trained language models, resulting in Boundary-Aware BERT (BABERT).We apply BABERT for feature induction of Chinese sequence labeling tasks.Experimental results on ten benchmarks of Chinese sequence labeling demonstrate that BABERT can provide consistent improvements on all datasets.In addition, our method can complement previous supervised lexicon exploration, where further improvements can be achieved when integrated with external lexicon information.

Transfer learning is a simple and powerful method that can be used to boost model performance of low-resource neural machine translation (NMT).Existing transfer learning methods for NMT are static, which simply transfer knowledge from a parent model to a child model once via parameter initialization.In this paper, we propose a novel transfer learning method for NMT, namely ConsistTL, which can continuously transfer knowledge from the parent model during the training of the child model.Specifically, for each training instance of the child model, ConsistTL constructs the semantically-equivalent instance for the parent model and encourages prediction consistency between the parent and child for this instance, which is equivalent to the child model learning each instance under the guidance of the parent model.Experimental results on five low-resource NMT tasks demonstrate that ConsistTL results in significant improvements over strong transfer learning baselines, with a gain up to 1.7 BLEU over the existing backtranslation model on the widely-used WMT17 Turkish-English benchmark.Further analysis reveals that ConsistTL can improve the inference calibration of the child model.Code and scripts are freely available at https://github. com/NLP2CT/ConsistTL.

Transformer-based autoregressive and nonautoregressive models have played an essential role in sequence generation tasks.The autoregressive model can obtain excellent performance, while the non-autoregressive model brings fast decoding speed for inference.In this paper, we propose JANUS, a Joint Autoregressive and Non-autoregressive training method using aUxiliary losS to enhance the model performance in both AR and NAR manner simultaneously and effectively alleviate the problem of distribution discrepancy.Further, we pre-train BART with JANUS on a large corpus with minimal cost (16 GPU days) and make the BART-JANUS capable of nonautoregressive generation, demonstrating that our approach can transfer the AR knowledge to NAR.Empirically, we show our approach and BART-JANUS can achieve significant improvement on multiple generation tasks, including machine translation and GLGE benchmarks.Our code is available at Github 1 .

Conventional phrase grounding aims to localize noun phrases mentioned in a given caption to their corresponding image regions, which has achieved great success recently.Apparently, sole noun phrase grounding is not enough for cross-modal visual language understanding.Here we extend the task by considering pronouns as well.First, we construct a dataset of phrase grounding with both noun phrases and pronouns to image regions.Based on the dataset, we test the performance of phrase grounding by using a state-of-the-art literature model of this line.Then, we enhance the baseline grounding model with coreference information which should help our task potentially, modeling the coreference structures with graph convolutional networks.Experiments on our dataset, interestingly, show that pronouns are easier to ground than noun phrases, where the possible reason might be that these pronouns are much less ambiguous.Additionally, our final model with coreference information can significantly boost the grounding performance of both noun phrases and pronouns.

In this paper we aim to relieve the issue of lexical translation inconsistency for documentlevel neural machine translation (NMT) by modeling consistency preference for lexical chains which consist of repeated words in a source-side document and provide a representation of the lexical consistency structure of the document.Specifically, we first propose lexical-consistency attention to capture consistency context among words in the same lexical chains.Then for each lexical chain we define and learn a consistency-tailored latent variable, which will guide the translation of corresponding sentences to enhance lexical translation consistency.Experimental results on Chinese→English and French→English document-level translation tasks show that our approach not only significantly improves translation performance in BLEU, but also substantially alleviates the problem of the lexical translation inconsistency.

Recent research has revealed that neural language models at scale suffer from poor temporal generalization capability, i.e., language model pre-trained on static data from past years performs worse over time on emerging data.Existing methods mainly perform continual training to mitigate such a misalignment.While effective to some extent but is far from being addressed on both the language modeling and downstream tasks.In this paper, we empirically observe that temporal generalization is closely affiliated with lexical semantic change, which is one of the essential phenomena of natural languages.Based on this observation, we propose a simple yet effective lexical-level masking strategy to post-train a converged language model.Experiments on two pre-trained language models, two different classification tasks, and four benchmark datasets demonstrate the effectiveness of our proposed method over existing temporal adaptation methods, i.e., continual training with new data.Our code is available at https: //github.com/zhaochen0110/LMLM.

Existing research generally treats Chinese character as a minimum unit for representation.However, such Chinese character representation will suffer two bottlenecks: 1) Learning bottleneck, the learning cannot benefit from its rich internal features (e.g., radicals and strokes); and 2) Parameter bottleneck, each individual character has to be represented by a unique vector.In this paper, we introduce a novel representation method for Chinese characters to break the bottlenecks, namely StrokeNet, which represents a Chinese character by a Latinized stroke sequence (e.g., "凹(concave)" to "ajaie" and "凸(convex)" to "aeaqe").Specifically, StrokeNet maps each stroke to a specific Latin character, thus allowing similar Chinese characters to have similar Latin representations.With the introduction of StrokeNet to neural machine translation (NMT), many powerful but not applicable techniques to non-Latin languages (e.g., shared subword vocabulary learning and ciphertextbased data augmentation) can now be perfectly implemented.Experiments on the widelyused NIST Chinese-English, WMT17 Chinese-English and IWSLT17 Japanese-English NMT tasks show that StrokeNet can provide a significant performance boost over the strong baselines with fewer model parameters, achieving 26.5 BLEU on the WMT17 Chinese-English task which is better than any previously reported results without using monolingual data.

Keyphrase generation aims to automatically generate short phrases summarizing an input document.The recently emerged ONE2SET paradigm (Ye et al., 2021) generates keyphrases as a set and has achieved competitive performance.Nevertheless, we observe serious calibration errors outputted by ONE2SET, especially in the over-estimation of ∅ token (means "no corresponding keyphrase").In this paper, we deeply analyze this limitation and identify two main reasons behind: 1) the parallel generation has to introduce excessive ∅ as padding tokens into training instances; and 2) the training mechanism assigning target to each slot is unstable and further aggravates the ∅ token over-estimation.To make the model wellcalibrated, we propose WR-ONE2SET which extends ONE2SET with an adaptive instancelevel cost Weighting strategy and a target Reassignment mechanism.The former dynamically penalizes the over-estimated slots for different instances thus smoothing the uneven training distribution.The latter refines the original inappropriate assignment and reduces the supervisory signals of over-estimated slots.Experimental results on commonly-used datasets demonstrate the effectiveness and generality of our proposed paradigm.

Image-to-text tasks, such as open-ended image captioning and controllable image description, have received extensive attention for decades.Here, we further advance this line of work by presenting Visual Spatial Description (VSD), a new perspective for image-to-text toward spatial semantics.Given an image and two objects inside it, VSD aims to produce one description focusing on the spatial perspective between the two objects.Accordingly, we manually annotate a dataset to facilitate the investigation of the newly-introduced task and build several benchmark encoder-decoder models by using VL-BART and VL-T5 as backbones.In addition, we investigate pipeline and joint end-to-end architectures for incorporating visual spatial relationship classification (VSRC) information into our model.Finally, we conduct experiments on our benchmark dataset to evaluate all our models.Results show that our models are impressive, providing accurate and human-like spatial-oriented text descriptions.Meanwhile, VSRC has great potential for VSD, and the joint end-to-end architecture is the better choice for their integration.We make the dataset and codes public for research purposes.

Text-based person search aims to retrieve images of a certain pedestrian by a textual description. The key challenge of this task is to eliminate the inter-modality gap and achieve the feature alignment across modalities. In this paper, we propose a semantic-aligned embedding method for text-based person search, in which the feature alignment across modalities is achieved by automatically learning the semantic-aligned visual features and textual features. First, we introduce two Transformer-based backbones to encode robust feature representations of the images and texts. Second, we design a semantic-aligned feature aggregation network to adaptively select and aggregate features with the same semantics into part-aware features, which is achieved by a multi-head attention module constrained by a cross-modality part alignment loss and a diversity loss. Experimental results on the CUHK-PEDES and Flickr30K datasets show that our method achieves state-of-the-art performances.

Most previous studies of document-level event extraction mainly focus on building argument chains in an autoregressive way, which achieves a certain success but is inefficient in both training and inference. In contrast to the previous studies, we propose a fast and lightweight model named as PTPCG. In our model, we design a novel strategy for event argument combination together with a non-autoregressive decoding algorithm via pruned complete graphs, which are constructed under the guidance of the automatically selected pseudo triggers. Compared to the previous systems, our system achieves competitive results with 19.8% of parameters and much lower resource consumption, taking only 3.8% GPU hours for training and up to 8.5 times faster for inference. Besides, our model shows superior compatibility for the datasets with (or without) triggers and the pseudo triggers can be the supplements for annotated triggers to make further improvements. Codes are available at https://github.com/Spico197/DocEE .

In the past few years, cross-modal image-text retrieval (ITR) has experienced increased interest in the research community due to its excellent research value and broad real-world application. It is designed for the scenarios where the queries are from one modality and the retrieval galleries from another modality. This paper presents a comprehensive and up-to-date survey on the ITR approaches from four perspectives. By dissecting an ITR system into two processes: feature extraction and feature alignment, we summarize the recent advance of the ITR approaches from these two perspectives. On top of this, the efficiency-focused study on the ITR system is introduced as the third perspective. To keep pace with the times, we also provide a pioneering overview of the cross-modal pre-training ITR approaches as the fourth perspective. Finally, we outline the common benchmark datasets and evaluation metric for ITR, and conduct the accuracy comparison among the representative ITR approaches. Some critical yet less studied issues are discussed at the end of the paper.

Document-level neural machine translation (DocNMT) universally encodes several local sentences or the entire document. Thus, DocNMT does not consider the relevance of document-level contextual information, for example, some context (i.e., content words, logical order, and co-occurrence relation) is more effective than another auxiliary context (i.e., functional and auxiliary words). To address this issue, we first utilize the word frequency information to recognize content words in the input document, and then use heuristical relations to summarize content words and sentences as a graph structure without relying on external syntactic knowledge. Furthermore, we apply graph attention networks to this graph structure to learn its feature representation, which allows DocNMT to more effectively capture the document-level context. Experimental results on several widely-used document-level benchmarks demonstrated the effectiveness of the proposed approach.

We propose a novel evaluation method for Question Generation (QG) task. It is designed to verify the quality of the generated questions in terms of different references, including not only the manually-written questions (i.e., ground truth) but also their variants. Back translation is utilized to obtain the variants, and accordingly, they generally appear as paraphrases of the ground-truth examples. In particular, an Asymmetrical Twin Gain (ATG) is proposed for binary-perspective evaluation using the existing metrics, such as BLEU and ROUGE-L, respectively. It enables both the metrics to be observed from two perspectives, including the consistency between QG results and ground-truth examples, as well as that of variants. The experiments on the publicly-available benchmark SQuAD demonstrate the reliability of ATG. More importantly, ATG is proven effective for indicating the stable QG performance. It is noteworthy that the proposed binary-perspective evaluation is explored for assisting the conventional evaluation methods, instead of replacing them. The contribute can be identified as the additional insight into the robustness of QG when some slightly-different references (e.g., paraphrases) are offered for evaluation. All the models and source codes in the experiments will be made publicly available to support reproducible research.

Medical dialogue generation is an important yet challenging task. Most previous works rely on the attention mechanism and large-scale pretrained language models. However, these methods often fail to acquire pivotal information from the long dialogue history to yield an accurate and informative response, due to the fact that the medical entities usually scatters throughout multiple utterances along with the complex relationships between them. To mitigate this problem, we propose a medical response generation model with Pivotal Information Recalling (MedPIR), which is built on two components, i.e., knowledge-aware dialogue graph encoder and recall-enhanced generator. The knowledge-aware dialogue graph encoder constructs a dialogue graph by exploiting the knowledge relationships between entities in the utterances, and encodes it with a graph attention network. Then, the recall-enhanced generator strengthens the usage of these pivotal information by generating a summary of the dialogue before producing the actual response. Experimental results on two large-scale medical dialogue datasets show that MedPIR outperforms the strong baselines in BLEU scores and medical entities F1 measure.

Pioneer efforts have been dedicated to the content-oriented video captioning that generates relevant sentences to describe the visual contents of a given video from the producer perspective. By contrast, this work targets at the search-oriented one that summarizes the given video via generating query-like sentences from the consumer angle. Beyond relevance, diversity is vital in characterizing consumers' seeking intention from different aspects. Towards this end, we devise a large-scale multimodal pre-training network regularized by five tasks to strengthen the downstream video representation, which is well-trained over our collected 11M micro-videos. Thereafter, we present a flow-based diverse captioning model to generate different captions from consumers' search demand. This model is optimized via a reconstruction loss and a KL divergence between the prior and the posterior. We justify our model over our constructed golden dataset comprising 690k pairs and experimental results demonstrate its superiority.

Visual Entailment with natural language explanations aims to infer the relationship between a text-image pair and generate a sentence to explain the decision-making process. Previous methods rely mainly on a pre-trained vision-language model to perform the relation inference and a language model to generate the corresponding explanation. However, the pre-trained vision-language models mainly build token-level alignment between text and image yet ignore the high-level semantic alignment between the phrases (chunks) and visual contents, which is critical for vision-language reasoning. Moreover, the explanation generator based only on the encoded joint representation does not explicitly consider the critical decision-making points of relation inference. Thus the generated explanations are less faithful to visual-language reasoning. To mitigate these problems, we propose a unified Chunk-aware Alignment and Lexical Constraint based method, dubbed as CALeC. It contains a Chunk-aware Semantic Interactor (arr. CSI), a relation inferrer, and a Lexical Constraint-aware Generator (arr. LeCG). Specifically, CSI exploits the sentence structure inherent in language and various image regions to build chunk-aware semantic alignment. Relation inferrer uses an attention-based reasoning network to incorporate the token-level and chunk-level vision-language representations. LeCG utilizes lexical constraints to expressly incorporate the words or chunks focused by the relation inferrer into explanation generation, improving the faithfulness and informativeness of the explanations. We conduct extensive experiments on three datasets, and experimental results indicate that CALeC significantly outperforms other competitor models on inference accuracy and quality of generated explanations.

Yue Zhang, Zhenghua Li, Zuyi Bao, Jiacheng Li, Bo Zhang, Chen Li, Fei Huang, Min Zhang. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

Yahui Liu, Haoping Yang, Chen Gong, Qingrong Xia, Zhenghua Li, Min Zhang. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

Linzhi Wu, Pengjun Xie, Jie Zhou, Meishan Zhang, Ma Chunping, Guangwei Xu, Min Zhang. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

Universally modeling all typical information extraction tasks (UIE) with one generative language model (GLM) has revealed great potential by the latest study, where various IE predictions are unified into a linearized hierarchical expression under a GLM. Syntactic structure information, a type of effective feature which has been extensively utilized in IE community, should also be beneficial to UIE. In this work, we propose a novel structure-aware GLM, fully unleashing the power of syntactic knowledge for UIE. A heterogeneous structure inductor is explored to unsupervisedly induce rich heterogeneous structural representations by post-training an existing GLM. In particular, a structural broadcaster is devised to compact various latent trees into explicit high-order forests, helping to guide a better generation during decoding. We finally introduce a task-oriented structure fine-tuning mechanism, further adjusting the learned structures to most coincide with the end-task's need. Over 12 IE benchmarks across 7 tasks our system shows significant improvements over the baseline UIE system. Further in-depth analyses show that our GLM learns rich task-adaptive structural bias that greatly resolves the UIE crux, the long-range dependence issue and boundary identifying.

How to effectively model global context has been a critical challenge for document-level neural machine translation (NMT). Both preceding and global context have been carefully explored in the sequence-to-sequence (seq2seq) framework. However, previous studies generally map global context into one vector, which is not enough to well represent the entire document since this largely ignores the hierarchy between sentences and words within. In this article, we propose to model global context for source language from both sentence level and word level. Specifically at sentence level, we extract useful global context for the current sentence, while at word level, we compute global context against words within the current sentence. On this basis, both kinds of global context can be appropriately fused before being incorporated into the state-of-the-art seq2seq model, i.e., Transformer . Detailed experimentation on various document-level translation tasks shows that global context at both sentence level and word level significantly improve translation performance. More encouraging, both kinds of global context are complementary. This leads to more improvement when both kinds of global context are used.

Recently, character information has been successfully introduced into the encoder-decoder event detection model to relieve the trigger-word mismatch problem, thus achieving impressive results in the languages without natural delimiters (i.e., Chinese). However, it is introduced into the encoder or the decoder separately, which makes the advantage of character information not be captured and represented adequately for event detection. In this article, we proposed a novel method to model character information in both the encoding and decoding stages to advance the neural event detection model. In particular, the proposed method can encode both words and characters and predict their event types jointly and further leverage interactions between word and its characters to optimize the inference. Experimental results show that the proposed model outperforms previous event detection methods on the ACE2005 Chinese benchmark. We release our code at Github. 1

Supervised statistical models rely on large-scale high-quality labeled data, which is important for model training but expensive to construct. Therefore, instead of constructing new dataset, researchers have attempted to make full use of various existing heterogeneous datasets to boost model performance, considering it is ubiquitous that the same task may have multiple annotated data following different and incompatible annotation guidelines. Representative methods include the guide-feature method which use the knowledge projected from the source-side to the target-side as extra features for target model guidance, and the multi-task learning (MTL) method which simultaneously train on multiple heterogeneous annotations with shared parameters to gain resource-share knowledge. Though effective, the guide-feature method fails to directly use the source-side data as training data, and the MTL method ignores the implicit mappings between heterogeneous datasets. Compared with the above methods, directly converting the heterogeneous datasets into homogeneous datasets for target model training is a more straightforward and effective way to fully exploit heterogeneous resources. In this work, we propose a neural coupled sequence labeling model for heterogeneous annotation conversion. First, for each token, we map a given one-side tag into a set of bundled tags by concatenating the tag with all the possible tags at the other side. Then, we build a neural coupled model over the bundled tag space. Finally, we convert heterogeneous annotations into homogeneous annotations by performing constraint decoding on the coupled model. We also propose a pruning strategy to address the oversize issue of the bundled tag space, which improves efficiency without hurting model performance.Experiments for part-of-speech (POS) tagging, word segmentation (WS), and WS&POS tagging tasks show that our proposed neural coupled model consistently outperforms several benchmark models for all the three tasks by large margin.

Knowledge distillation (KD) is an effective strategy for neural machine translation (NMT) to improve the performance of a student model. Usually, the teacher can guide the student to be better by distilling the soft label or data knowledge from the teacher itself. However, the data diversity and teacher knowledge are limited with only one teacher model. Though a natural solution is to adopt multiple randomized teacher models, one big shortcoming is that the model parameters and training costs are largely increased with the number of teacher models. In this work, we explore to mimic multiple teacher distillation from the sub-network space and permuted variants of one single teacher model. Specifically, we train a teacher by multiple sub-network extraction paradigms: sub-layer reordering, layer-drop, and dropout variants. In doing so, one teacher model can provide multiple outputs variants and causes neither additional parameters nor much extra training cost. Experiments on $8$ IWSLT datasets: (IWSLT14 En $\leftrightarrow$ De, En $\leftrightarrow$ Es, and IWSLT17 En $\leftrightarrow$ Fr, En $\leftrightarrow$ Zh) and the large WMT14 EN $\to$ DE translation tasks show that our method even achieves nearly comparable performance with multiple teacher models with different randomized parameters, both word-level, and sequence-level knowledge distillation. Our code is available at GitHub\footnote{https://github.com/dropreg/RLD}

Existing studies for multi-source neural machine translation (NMT) either separately model different source sentences or resort to the conventional single-source NMT by simply concatenating all source sentences. However, there exist two drawbacks in these approaches. First, they ignore the explicit word-level semantic interactions between source sentences, which have been shown effective in the embeddings of multilingual texts. Second, multiple source sentences are simultaneously encoded by an NMT model, which is unable to fully exploit the semantic information of each source sentence. In this paper, we explore multi-stage information interactions for multi-source NMT. Specifically, we first propose a multi-source NMT model that performs information interactions at the encoding stage. Its encoder contains multiple semantic interaction layers, each of which sequentially consists of (1) monolingual semantic interaction sub-layer, which is based on the self-attention mechanism and used to learn word-level monolingual contextual representations of source sentences, and (2) cross-lingual semantic interaction sub-layer, which leverages word alignments to perform fine-grained semantic transitions among hidden states of different source sentences. Furthermore, at the training stage, we introduce a mutual distillation based training framework, where single-source models and ours perform information interactions. Such framework can fully exploit the semantic information of each source sentence to enhance our model. Extensive experimental results on the WMT14 English-German-French dataset show our method exhibits significant improvements upon competitive baselines.

The encoder–decoder framework has been widely used in various practical artificial intelligence cyber-physical systems, including intelligent translation systems. The decoding process in such a framework usually demands the target-side representation, which is often learned by an autoaggressive decoder to simulate the target context information at the current time-step. However, the autoaggressive decoder only captures the previously generated partial target fragment and fails in simulating the global contextual information. In this article, we propose a new data-driven fuzzy context representation strategy to simulate the global target information. Specifically, we design two fuzzy methods to the global target contextual information, which are bag-of-words of target language generated via a softmax layer from the source-side representation and whole target sentence retrieved from the translation memory according to the source-side representation. Both methods facilitate the autoaggressive decoder to handle the global target context at the current time-step, thereby learning a more effective context vector for the generation of target translation. Extensive experiments on two machine translation tasks demonstrated that the proposed method achieved 3% improvement of BLEU score over a strong baseline.

Recent studies have pointed out that many well-developed Visual Question Answering (VQA) models are heavily affected by the language prior problem. It refers to making predictions based on the co-occurrence pattern between textual questions and answers instead of reasoning upon visual contents. To tackle this problem, most existing methods focus on strengthening the visual feature learning capability to reduce this text shortcut influence on model decisions. However, few efforts have been devoted to analyzing its inherent cause and providing an explicit interpretation. It thus lacks a good guidance for the research community to move forward in a purposeful way, resulting in model construction perplexity towards overcoming this non-trivial problem. In this paper, we propose to interpret the language prior problem in VQA from a class-imbalance view. Concretely, we design a novel interpretation scheme whereby the loss of mis-predicted frequent and sparse answers from the same question type is distinctly exhibited during the late training phase. It explicitly reveals why the VQA model tends to produce a frequent yet obviously wrong answer, to a given question whose right answer is sparse in the training set. Based upon this observation, we further propose a novel loss re-scaling approach to assign different weights to each answer according to the training data statistics for estimating the final loss. We apply our approach into six strong baselines and the experimental results on two VQA-CP benchmark datasets evidently demonstrate its effectiveness. In addition, we also justify the validity of the class imbalance interpretation scheme on other computer vision tasks, such as face recognition and image classification.

Visual attention in Visual Question Answering (VQA) targets at locating the right image regions regarding the answer prediction, offering a powerful technique to promote multi-modal understanding. However, recent studies have pointed out that the highlighted image regions from the visual attention are often irrelevant to the given question and answer, leading to model confusion for correct visual reasoning. To tackle this problem, existing methods mostly resort to aligning the visual attention weights with human attentions. Nevertheless, gathering such human data is laborious and expensive, making it burdensome to adapt well-developed models across datasets. To address this issue, in this article, we devise a novel visual attention regularization approach, namely, AttReg, for better visual grounding in VQA. Specifically, AttReg first identifies the image regions that are essential for question answering yet unexpectedly ignored (i.e., assigned with low attention weights) by the backbone model. And then a mask-guided learning scheme is leveraged to regularize the visual attention to focus more on these ignored key regions. The proposed method is very flexible and model-agnostic, which can be integrated into most visual attention-based VQA models and require no human attention supervision. Extensive experiments over three benchmark datasets, i.e., VQA-CP v2, VQA-CP v1, and VQA v2, have been conducted to evaluate the effectiveness of AttReg. As a by-product, when incorporating AttReg into the strong baseline LMH, our approach can achieve a new state-of-the-art accuracy of 60.00% with an absolute performance gain of 7.01% on the VQA-CP v2 benchmark dataset. In addition to the effectiveness validation, we recognize that the faithfulness of the visual attention in VQA has not been well explored in literature. In the light of this, we propose to empirically validate such property of visual attention and compare it with the prevalent gradient-based approaches.

Linqing Chen, Junhui Li, Zhengxian Gong, Boxing Chen, Weihua Luo, Min Zhang, Guodong Zhou. 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.

Chen Gong, Saihao Huang, Houquan Zhou, Zhenghua Li, Min Zhang, Zhefeng Wang, Baoxing Huai, Nicholas Jing Yuan. 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.

Xin Liu, Baosong Yang, Dayiheng Liu, Haibo Zhang, Weihua Luo, Min Zhang, Haiying Zhang, Jinsong Su. 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.

Dongqin Xu, Junhui Li, Muhua Zhu, Min Zhang, Guodong Zhou. 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.

Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou. 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.

The most straightforward approach to joint word segmentation (WS), part-of-speech (POS) tagging, and constituent parsing (PAR) is converting a word-level tree into a char-level tree, which, however, leads to two severe challenges.First, a larger label set (e.g., ≥ 600) and longer inputs both increase computational cost.Second, it is difficult to rule out illegal trees containing conflicting production rules, which is important for reliable model evaluation.If a POS tag (like VV) is above a phrase tag (like VP) in the output tree, it becomes quite complex to decide word boundaries.To deal with both challenges, this work proposes a two-stage coarse-to-fine labeling framework for joint WS-POS-PAR.In the coarse labeling stage, the joint model outputs a bracketed tree, in which each node corresponds to one of four labels (i.e., phrase, subphrase, word, subword).The tree is guaranteed to be legal via constrained CKY decoding.In the fine labeling stage, the model expands each coarse label into a final label (such as VP, VP * , VV, VV * ).Experiments on Chinese Penn Treebank 5.1 and 7.0 show that our joint model consistently outperforms the pipeline approach on both settings of without and with BERT, and achieves new state-of-the-art performance.

Aspect terms extraction (ATE) and aspect sentiment classification (ASC) are two fundamental and fine-grained sub-tasks in aspect-level sentiment analysis (ALSA).In the textual analysis, jointly extracting both aspect terms and sentiment polarities has been drawn much attention due to the better applications than individual sub-task.However, in the multimodal scenario, the existing studies are limited to handle each sub-task independently, which fails to model the innate connection between the above two objectives and ignores the better applications.Therefore, in this paper, we are the first to jointly perform multi-modal ATE (MATE) and multi-modal ASC (MASC), and we propose a multi-modal joint learning approach with auxiliary cross-modal relation detection for multi-modal aspect-level sentiment analysis (MALSA).Specifically, we first build an auxiliary text-image relation detection module to control the proper exploitation of visual information.Second, we adopt the hierarchical framework to bridge the multi-modal connection between MATE and MASC, as well as separately visual guiding for each sub module.Finally, we can obtain all aspect-level sentiment polarities dependent on the jointly extracted specific aspects.Extensive experiments show the effectiveness of our approach against the joint textual approaches, pipeline and collapsed multi-modal approaches.

Recently a number of approaches have been proposed to improve translation performance for document-level neural machine translation (NMT).However, few are focusing on the subject of lexical translation consistency.In this paper we apply "one translation per discourse" in NMT, and aim to encourage lexical translation consistency for document-level NMT.This is done by first obtaining a word link for each source word in a document, which tells the positions where the source word appears at.Then we encourage the translations of those words within a link to be consistent in two ways.On the one hand, when encoding sentences within a document we properly exchange context information of those words.On the other hand, we propose an auxiliary loss function to better constrain that their translations should be consistent.Experimental results on Chinese↔English and English→French translation tasks show that our approach not only achieves state-of-the-art performance in BLEU scores, but also greatly improves lexical translation consistency.

Data augmentation has attracted a lot of research attention in the deep learning era for its ability in alleviating data sparseness.The lack of labeled data for unseen evaluation databases is exactly the major challenge for cross-domain text-to-SQL parsing.Previous works either require human intervention to guarantee the quality of generated data, or fail to handle complex SQL queries.This paper presents a simple yet effective data augmentation framework.First, given a database, we automatically produce a large number of SQL queries based on an abstract syntax tree grammar.For better distribution matching, we require that at least 80% of SQL patterns in the training data are covered by generated queries.Second, we propose a hierarchical SQL-to-question generation model to obtain high-quality natural language questions, which is the major contribution of this work.Finally, we design a simple sampling strategy that can greatly improve training efficiency given large amounts of generated data.Experiments on three cross-domain datasets, i.e., WikiSQL and Spider in English, and DuSQL in Chinese, show that our proposed data augmentation framework can consistently improve performance over strong baselines, and the hierarchical generation component is the key for the improvement.

Document context-aware machine translation remains challenging due to the lack of large-scale document parallel corpora. To make full use of source-side monolingual documents for context-aware NMT, we propose a Pre-training approach with Global Context (PGC). In particular, we first propose a novel self-supervised pre-training task, which contains two training objectives: (1) reconstructing the original sentence from a corrupted version; (2) generating a gap sentence from its left and right neighbouring sentences. Then we design a universal model for PGC which consists of a global context encoder, a sentence encoder and a decoder, with similar architecture to typical context-aware NMT models. We evaluate the effectiveness and generality of our pre-trained PGC model by adapting it to various downstream context-aware NMT models. Detailed experimentation on four different translation tasks demonstrates that our PGC approach significantly improves the translation performance of context-aware NMT. For example, based on the state-of-the-art SAN model, we achieve an averaged improvement of 1.85 BLEU scores and 1.59 Meteor scores on the four translation tasks.

Qingrong Xia, Bo Zhang, Rui Wang, Zhenghua Li, Yue Zhang, Fei Huang, Luo Si, Min Zhang. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

Dropout is a powerful and widely used technique to regularize the training of deep neural networks. Though effective and performing well, the randomness introduced by dropout causes unnegligible inconsistency between training and inference. In this paper, we introduce a simple consistency training strategy to regularize dropout, namely R-Drop, which forces the output distributions of different sub models generated by dropout to be consistent with each other. Specifically, for each training sample, R-Drop minimizes the bidirectional KL-divergence between the output distributions of two sub models sampled by dropout. Theoretical analysis reveals that R-Drop reduces the above inconsistency. Experiments on $\bf{5}$ widely used deep learning tasks ($\bf{18}$ datasets in total), including neural machine translation, abstractive summarization, language understanding, language modeling, and image classification, show that R-Drop is universally effective. In particular, it yields substantial improvements when applied to fine-tune large-scale pre-trained models, e.g., ViT, RoBERTa-large, and BART, and achieves state-of-the-art (SOTA) performances with the vanilla Transformer model on WMT14 English$\to$German translation ($\bf{30.91}$ BLEU) and WMT14 English$\to$French translation ($\bf{43.95}$ BLEU), even surpassing models trained with extra large-scale data and expert-designed advanced variants of Transformer models. Our code is available at GitHub\footnote{\url{https://github.com/dropreg/R-Drop}}.

Existing multi-turn context-response matching methods mainly concentrate on obtaining multi-level and multi-dimension representations and better interactions between context utterances and response. However, in real-place conversation scenarios, whether a response candidate is suitable not only counts on the given dialogue context but also other backgrounds, e.g., wording habits, user-specific dialogue history content. To fill the gap between these up-to-date methods and the real-world applications, we incorporate user-specific dialogue history into the response selection and propose a personalized hybrid matching network (PHMN). Our contributions are two-fold: (1) our model extracts personalized wording behaviors from user-specific dialogue history as extra matching information; (2) we perform hybrid representation learning on context-response utterances and explicitly incorporate a customized attention mechanism to extract vital information from context-response interactions so as to improve the accuracy of matching. We evaluate our model on two large datasets with user identification, i.e., personalized Ubuntu dialogue Corpus (P-Ubuntu) and personalized Weibo dataset (P-Weibo). Experimental results confirm that our method significantly outperforms several strong models by combining personalized attention, wording behaviors, and hybrid representation learning.

We present a novel approach to improve the performance of distant supervision relation extraction with Positive and Unlabeled (PU) Learning. This approach first applies reinforcement learning to decide whether a sentence is positive to a given relation, and then positive and unlabeled bags are constructed. In contrast to most previous studies, which mainly use selected positive instances only, we make full use of unlabeled instances and propose two new representations for positive and unlabeled bags. These two representations are then combined in an appropriate way to make bag-level prediction. Experimental results on a widely used real-world dataset demonstrate that this new approach indeed achieves significant and consistent improvements as compared to several competitive baselines.

Transfer learning between different language pairs has shown its effectiveness for Neural Machine Translation (NMT) in low-resource scenario. However, existing transfer methods involving a common target language are far from success in the extreme scenario of zero-shot translation, due to the language space mismatch problem between transferor (the parent model) and transferee (the child model) on the source side. To address this challenge, we propose an effective transfer learning approach based on cross-lingual pre-training. Our key idea is to make all source languages share the same feature space and thus enable a smooth transition for zero-shot translation. To this end, we introduce one monolingual pre-training method and two bilingual pre-training methods to obtain a universal encoder for different languages. Once the universal encoder is constructed, the parent model built on such encoder is trained with large-scale annotated data and then directly applied in zero-shot translation scenario. Experiments on two public datasets show that our approach significantly outperforms strong pivot-based baseline and various multilingual NMT approaches.

We investigate the task of constraining NMT with pre-specified translations, which has practical significance for a number of research and industrial applications. Existing works impose pre-specified translations as lexical constraints during decoding, which are based on word alignments derived from target-to-source attention weights. However, multiple recent studies have found that word alignment derived from generic attention heads in the Transformer is unreliable. We address this problem by introducing a dedicated head in the multi-head Transformer architecture to capture external supervision signals. Results on five language pairs show that our method is highly effective in constraining NMT with pre-specified translations, consistently outperforming previous methods in translation quality.

Sentiment analysis in dialogues plays a critical role in dialogue data analysis. However, previous studies on sentiment classification in dialogues largely ignore topic information, which is important for capturing overall information in some types of dialogues. In this study, we focus on the sentiment classification task in an important type of dialogue, namely customer service dialogue, and propose a novel approach which captures overall information to enhance the classification performance. Specifically, we propose a topic-aware multi-task learning (TML) approach which learns topic-enriched utterance representations in customer service dialogue by capturing various kinds of topic information. In the experiment, we propose a large-scale and high-quality annotated corpus for the sentiment classification task in customer service dialogue and empirical studies on the proposed corpus show that our approach significantly outperforms several strong baselines.

In the literature, existing studies always consider Aspect Sentiment Classification (ASC) as an independent sentence-level classification problem aspect by aspect, which largely ignore the document-level sentiment preference information, though obviously such information is crucial for alleviating the information deficiency problem in ASC. In this paper, we explore two kinds of sentiment preference information inside a document, i.e., contextual sentiment consistency w.r.t. the same aspect (namely intra-aspect sentiment consistency) and contextual sentiment tendency w.r.t. all the related aspects (namely inter-aspect sentiment tendency). On the basis, we propose a Cooperative Graph Attention Networks (CoGAN) approach for cooperatively learning the aspect-related sentence representation. Specifically, two graph attention networks are leveraged to model above two kinds of document-level sentiment preference information respectively, followed by an interactive mechanism to integrate the two-fold preference. Detailed evaluation demonstrates the great advantage of the proposed approach to ASC over the state-of-the-art baselines. This justifies the importance of the document-level sentiment preference information to ASC and the effectiveness of our approach capturing such information.

International audience

In the deep learning (DL) era, parsing models are extremely simplified with little hurt on performance, thanks to the remarkable capability of multi-layer BiLSTMs in context representation.As the most popular graphbased dependency parser due to its high efficiency and performance, the biaffine parser directly scores single dependencies under the arc-factorization assumption, and adopts a very simple local token-wise cross-entropy training loss.This paper for the first time presents a second-order TreeCRF extension to the biaffine parser.For a long time, the complexity and inefficiency of the inside-outside algorithm hinder the popularity of TreeCRF.To address this issue, we propose an effective way to batchify the inside and Viterbi algorithms for direct large matrix operation on GPUs, and to avoid the complex outside algorithm via efficient back-propagation.Experiments and analysis on 27 datasets from 13 languages clearly show that techniques developed before the DL era, such as structural learning (global TreeCRF loss) and high-order modeling are still useful, and can further boost parsing performance over the state-of-the-art biaffine parser, especially for partially annotated training data.We release our code at https: //github.com/yzhangcs/crfpar.

Opinion role labeling (ORL) is a fine-grained opinion analysis task and aims to answer "who expressed what kind of sentiment towards what?".Due to the scarcity of labeled data, ORL remains challenging for data-driven methods.In this work, we try to enhance neural ORL models with syntactic knowledge by comparing and integrating different representations.We also propose dependency graph convolutional networks (DEPGCN) to encode parser information at different processing levels.In order to compensate for parser inaccuracy and reduce error propagation, we introduce multi-task learning (MTL) to train the parser and the ORL model simultaneously.We verify our methods on the benchmark MPQA corpus.The experimental results show that syntactic information is highly valuable for ORL, and our final MTL model effectively boosts the F1 score by 9.29 over the syntaxagnostic baseline.In addition, we find that the contributions from syntactic knowledge do not fully overlap with contextualized word representations (BERT).Our best model achieves 4.34 higher F1 score than the current state-ofthe-art.

In contrast with the traditional single-grained word segmentation (SWS), where a sentence corresponds to a single word sequence, multi-grained Chinese word segmentation (MWS) aims to segment a sentence into multiple word sequences to preserve all words of different granularities.Due to the lack of manually annotated MWS data, previous work train and tune MWS models only on automatically generated pseudo MWS data.In this work, we further take advantage of the rich word boundary information in existing SWS data and naturally annotated data from dictionary example (DictEx) sentences, to advance the state-of-the-art MWS model based on the idea of weak supervision.Particularly, we propose to accommodate two types of weakly labeled data for MWS, i.e., SWS data and DictEx data by employing a simple yet competitive graph-based parser with local loss.Besides, we manually annotate a high-quality MWS dataset according to our newly compiled annotation guideline, consisting of over 9,000 sentences from two types of texts, i.e., canonical newswire (NEWS) and non-canonical web (BAIKE) data for better evaluation.Detailed evaluation shows that our proposed model with weakly labeled data significantly outperforms the state-of-the-art MWS model by 1.12 and 5.97 on NEWS and BAIKE data in F1.

In recent years, parsing performance is dramatically improved on in-domain texts thanks to the rapid progress of deep neural network models.The major challenge for current parsing research is to improve parsing performance on out-of-domain texts that are very different from the indomain training data when there is only a small-scale out-domain labeled data.To deal with this problem, we propose to improve the contextualized word representations via adversarial learning and fine-tuning BERT processes.Concretely, we apply adversarial learning to three representative semi-supervised domain adaption methods, i.e., direct concatenation (CON), feature augmentation (FA), and domain embedding (DE) with two useful strategies, i.e., fused targetdomain word representations and orthogonality constraints, thus enabling to model more pure yet effective domain-specific and domain-invariant representations.Simultaneously, we utilize a large-scale target-domain unlabeled data to fine-tune BERT with only the language model loss, thus obtaining reliable contextualized word representations that benefit for the cross-domain dependency parsing.Experiments on a benchmark dataset show that our proposed adversarial approaches achieve consistent improvements, and fine-tuning BERT further boosts the parsing accuracy by a large margin.Our single model achieves the same state-of-the-art performance as the top submitted system in the NLPCC-2019 shared task, which uses ensemble models and BERT.

We tackle implicit discourse relation recognition.Both self-attention and interactive-attention mechanisms have been applied for attention-aware representation learning, which improves the current discourse analysis models.To take advantages of the two attention mechanisms simultaneously, we develop a propagative attention learning model using a cross-coupled two-channel network.We experiment on Penn Discourse Treebank.The test results demonstrate that our model yields substantial improvements over the baselines (BiLSTM and BERT).

Recently, due to the interplay between syntax and semantics, incorporating syntactic knowledge into neural semantic role labeling (SRL) has achieved much attention.Most of the previous syntax-aware SRL works focus on explicitly modeling homogeneous syntactic knowledge over tree outputs.In this work, we propose to encode heterogeneous syntactic knowledge for SRL from both explicit and implicit representations.First, we introduce graph convolutional networks to explicitly encode multiple heterogeneous dependency parse trees.Second, we extract the implicit syntactic representations from syntactic parser trained with heterogeneous treebanks.Finally, we inject the two types of heterogeneous syntax-aware representations into the base SRL model as extra inputs.We conduct experiments on two widely-used benchmark datasets, i.e., Chinese Proposition Bank 1.0 and English CoNLL-2005 dataset.Experimental results show that incorporating heterogeneous syntactic knowledge brings significant improvements over strong baselines.We further conduct detailed analysis to gain insights on the usefulness of heterogeneous (vs.homogeneous) syntactic knowledge and the effectiveness of our proposed approaches for modeling such knowledge.

Supervised models for Relation Extraction (RE) typically require human-annotated training data.Due to the limited size, the human-annotated data is usually incapable of covering diverse relation expressions, which could limit the performance of RE.To increase the coverage of relation expressions, we may enlarge the labeled data by hiring annotators or applying Distant Supervision (DS).However, the human-annotated data is costly and non-scalable while the distantly supervised data contains many noises.In this paper, we propose an alternative approach to improve RE systems via enriching diverse expressions by relational paraphrase sentences.Based on an existing labeled data, we first automatically build a task-specific paraphrase data.Then, we propose a novel model to learn the information of diverse relation expressions.In our model, we try to capture this information on the paraphrases via a joint learning framework.Finally, we conduct experiments on a widely used dataset and the experimental results show that our approach is effective to improve the performance on relation extraction, even compared with a strong baseline.

Neural machine translation with millions of parameters is vulnerable to unfamiliar inputs.We propose Token Drop to improve generalization and avoid overfitting for the NMT model.Similar to word dropout, whereas we replace dropped token with a special token instead of setting zero to words.We further introduce two self-supervised objectives: Replaced Token Detection and Dropped Token Prediction.Our method aims to force model generating target translation with less information, in this way the model can learn textual representation better.Experiments on Chinese-English and English-Romanian benchmark demonstrate the effectiveness of our approach and our model achieves significant improvements over a strong Transformer baseline 1 .

In recent years, distantly-supervised relation extraction has achieved a certain success by using deep neural networks.Distant Supervision (DS) can automatically generate large-scale annotated data by aligning entity pairs from Knowledge Bases (KB) to sentences.However, these DSgenerated datasets inevitably have wrong labels that result in incorrect evaluation scores during testing, which may mislead the researchers.To solve this problem, we build a new dataset NYT-H, where we use the DS-generated data as training data and hire annotators to label test data.Compared with the previous datasets, NYT-H has a much larger test set and then we can perform more accurate and consistent evaluation.Finally, we present the experimental results of several widely used systems on NYT-H.The experimental results show that the ranking lists of the comparison systems on the DS-labelled test data and human-annotated test data are different.This indicates that our human-annotated data is necessary for evaluation of distantly-supervised relation extraction.

Due to the lack of labeled data, previous research on text-to-SQL parsing mainly focuses on English.Representative English datasets include ATIS, WikiSQL, Spider, etc.This paper presents DuSQL, a larges-scale and pragmatic Chinese dataset for the cross-domain text-to-SQL task, containing 200 databases, 813 tables, and 23,797 question/SQL pairs.Our new dataset has three major characteristics.First, by manually analyzing questions from several representative applications, we try to figure out the true distribution of SQL queries in real-life needs.Second, DuSQL contains a considerable proportion of SQL queries involving row or column calculations, motivated by our analysis on the SQL query distributions.Finally, we adopt an effective data construction framework via human-computer collaboration.The basic idea is automatically generating SQL queries based on the SQL grammar and constrained by the given database.This paper describes in detail the construction process and data statistics of DuSQL.Moreover, we present and compare performance of several open-source textto-SQL parsers with minor modification to accommodate Chinese, including a simple yet effective extension to IRNet for handling calculation SQL queries.

In the literature, the research on abstract meaning representation (AMR) parsing is much restricted by the size of human-curated dataset which is critical to build an AMR parser with good performance.To alleviate such data size restriction, pre-trained models have been drawing more and more attention in AMR parsing.However, previous pre-trained models, like BERT, are implemented for general purpose which may not work as expected for the specific task of AMR parsing.In this paper, we focus on sequence-to-sequence (seq2seq) AMR parsing and propose a seq2seq pre-training approach to build pre-trained models in both single and joint way on three relevant tasks, i.e., machine translation, syntactic parsing, and AMR parsing itself.Moreover, we extend the vanilla fine-tuning method to a multi-task learning fine-tuning method that optimizes for the performance of AMR parsing while endeavors to preserve the response of pre-trained models.Extensive experimental results on two English benchmark datasets show that both the single and joint pre-trained models significantly improve the performance (e.g., from 71.5 to 80.2 on AMR 2.0), which reaches the state of the art.The result is very encouraging since we achieve this with seq2seq models rather than complex models.We make our code and model available at https:// github.com/xdqkid/S2S-AMR-Parser.

In this article, we show that word translations can be explicitly incorporated into NMT effectively to avoid wrong translations. Specifically, we propose three cross-lingual encoders to explicitly incorporate word translations into NMT: (1) Factored encoder, which encodes a word and its translation in a vertical way; (2) Gated encoder, which uses a gated mechanism to selectively control the amount of word translations moving forward; and (3) Mixed encoder, which stitchingly learns a word and its translation annotations over sequences where words and their translations are alternatively mixed. Besides, we first use a simple word dictionary approach and then a word sense disambiguation (WSD) approach to effectively model the word context for better word translation. Experimentation on Chinese-to-English translation demonstrates that all proposed encoders are able to improve the translation accuracy for both traditional RNN-based NMT and recent self-attention-based NMT (hereafter referred to as Transformer ). Specifically, Mixed encoder yields the most significant improvement of 2.0 in BLEU on the RNN-based NMT, while Gated encoder improves 1.2 in BLEU on Transformer . This indicates the usefulness of an WSD approach in modeling word context for better word translation. This also indicates the effectiveness of our proposed cross-lingual encoders in explicitly modeling word translations to avoid wrong translations in NMT. Finally, we discuss in depth how word translations benefit different NMT frameworks from several perspectives.

Chinese zero pronoun (ZP) resolution plays a critical role in discourse analysis. Different from traditional mention-to-mention approaches, this article proposes a chain-to-chain approach to improve the performance of ZP resolution in three aspects. First, consecutive ZPs are clustered into coreferential chains, each working as one independent anaphor as a whole. In this way, those ZPs far away from their overt antecedents can be bridged via other consecutive ZPs in the same coreferential chains and thus better resolved. Second, common noun phrases (NPs) are automatically grouped into coreferential chains using traditional approaches, each working as one independent antecedent candidate as a whole. That is, those NPs occurring in the same coreferential chain are viewed as one antecedent candidate as a whole, and ZP resolution is made between ZP coreferential chains and common NP coreferential chains. In this way, the performance can be much improved due to the effective reduction of the search space by pruning singletons and negative instances. Third and finally, additional features from ZP and common NP coreferential chains are employed to better represent anaphors and their antecedent candidates, respectively. Comprehensive experiments on the OntoNotes V5.0 corpus show that our chain-to-chain approach significantly outperforms the state-of-the-art mention-to-mention approaches. To our knowledge, this is the first work to resolve zero pronouns in a chain-to-chain way.

In neural machine translation (NMT), the source and target words are at the two ends of a large deep neural network, normally mediated by a series of non-linear activations. The problem with such consequent non-linear activations is that they significantly decrease the magnitude of the gradient in a deep neural network, and thus gradually loosen the interaction between source words and their translations. As a result, a source word may be incorrectly translated into a target word out of its translational equivalents. In this article, we propose short-path units (SPUs) to strengthen the association of source and target words by allowing information flow over adjacent layers effectively via linear interpolation. In particular, we enrich three critical NMT components with SPUs: (1) an enriched encoding model with SPU, which interpolates source word embeddings linearly into source annotations; (2) an enriched decoding model with SPU, which enables the source context linearly flow to target-side hidden states; and (3) an enriched output model with SPU, which further allows linear interpolation of target-side hidden states into output states. Experimentation on Chinese-to-English, English-to-German, and low-resource Tibetan-to-Chinese translation tasks demonstrates that the linear interpolation of SPUs significantly improves the overall translation quality by 1.88, 1.43, and 3.75 BLEU, respectively. Moreover, detailed analysis shows that our approaches much strengthen the association of source and target words. From the preceding, we can see that our proposed model is effective both in rich- and low-resource scenarios.

The convolutional sequence-to-sequence (ConvS2S) machine translation system is one of the typical neural machine translation (NMT) systems. Training the ConvS2S model tends to get stuck in a local optimum in our pre-studies. To overcome this inferior behavior, we propose to de-train a trained ConvS2S model in a mild way and retrain to find a better solution globally. In particular, the trained parameters of one layer of the NMT network are abandoned by re-initialization while other layers’ parameters are kept at the same time to kick off re-optimization from a new start point and safeguard the new start point not too far from the previous optimum. This procedure is executed layer by layer until all layers of the ConvS2S model are explored. Experiments show that when compared to various measures for escaping from the local optimum, including initialization with random seeds, adding perturbations to the baseline parameters, and continuing training (con-training) with the baseline models, our method consistently improves the ConvS2S translation quality across various language pairs and achieves better performance.

While neural models based on the Transformer architecture achieve the State-of-the-Art translation performance, it is well known that the learned target-to-source attentions do not correlate well with word alignment. There is an increasing interest in inducing accurate word alignment in Transformer, due to its important role in practical applications such as dictionary-guided translation and interactive translation. In this article, we extend and improve the recent work on unsupervised learning of word alignment in Transformer on two dimensions: a) parameter initialization from a pre-trained cross-lingual language model to leverage large amounts of monolingual data for learning robust contextualized word representations, and b) regularization of the training objective to directly model characteristics of word alignments which results in favorable word alignments receiving more concentrated probabilities. Experiments on benchmark data sets of three language pairs show that the proposed methods can significantly reduce alignment error rate (AER) by at least 3.7 to 7.7 points on each language pair over two recent works on improving the Transformer's word alignment. Moreover, our methods can achieve better alignment results than GIZA++ on certain test sets.

In neural machine translation (NMT), there is a natural correspondence between source and target sentences. The traditional NMT method does not explicitly model the translation agreement on sentence-level. In this article, we propose a comprehensive and novel sentence-level agreement architecture to alleviate this problem. It directly minimizes the difference between the representations of the source-side and target-side sentence on sentence-level. First, we compare a variety of sentence representation strategies and propose a “Gated Sum” sentence representation to achieve better sentence semantic information. Then, rather than a single-layer sentence-level agreement architecture, we further propose a multi-layer sentence agreement architecture to make the source and target semantic spaces closer layer by layer. The proposed agreement module can be integrated into NMT as an additional training objective function, and can also be used to enhance the representation of the source-side sentences. Experiments on the NIST Chinese-to-English and the WMT English-to-German translation tasks show that the proposed agreement architecture achieves significant improvements over state-of-the-art baselines, demonstrating the effectiveness and necessity of exploiting sentence-level agreement for NMT.

The graph-based semisupervised label propagation (LP) algorithm has delivered impressive classification results. However, the estimated soft labels typically contain mixed signs and noise, which cause inaccurate predictions due to the lack of suitable constraints. Moreover, the available methods typically calculate the weights and estimate the labels in the original input space, which typically contains noise and corruption. Thus, the encoded similarities and manifold smoothness may be inaccurate for label estimation. In this article, we present effective schemes for resolving these issues and propose a novel and robust semisupervised classification algorithm, namely the triple matrix recovery-based robust auto-weighted label propagation framework (ALP-TMR). Our ALP-TMR introduces a TMR mechanism to remove noise or mixed signs from the estimated soft labels and improve the robustness to noise and outliers in the steps of assigning weights and predicting the labels simultaneously. Our method can jointly recover the underlying clean data, clean labels, and clean weighting spaces by decomposing the original data, predicted soft labels, or weights into a clean part plus an error part by fitting noise. In addition, ALP-TMR integrates the auto-weighting process by minimizing the reconstruction errors over the recovered clean data and clean soft labels, which can encode the weights more accurately to improve both data representation and classification. By classifying samples in the recovered clean label and weight spaces, one can potentially improve the label prediction results. Extensive simulations verified the effectivenss of our ALP-TMR.

Semantic role labeling (SRL), also known as shallow semantic parsing, is an important yet challenging task in NLP. Motivated by the close correlation between syntactic and semantic structures, traditional discrete-feature-based SRL approaches make heavy use of syntactic features. In contrast, deep-neural-network-based approaches usually encode the input sentence as a word sequence without considering the syntactic structures. In this work, we investigate several previous approaches for encoding syntactic trees, and make a thorough study on whether extra syntax-aware representations are beneficial for neural SRL models. Experiments on the benchmark CoNLL-2005 dataset show that syntax-aware SRL approaches can effectively improve performance over a strong baseline with external word representations from ELMo. With the extra syntax-aware representations, our approaches achieve new state-of-the-art 85.6 F1 (single model) and 86.6 F1 (ensemble) on the test data, outperforming the corresponding strong baselines with ELMo by 0.8 and 1.0, respectively. Detailed error analysis are conducted to gain more insights on the investigated approaches.

Abstractive Sentence Summarization (AS-SUM) targets at grasping the core idea of the source sentence and presenting it as the summary.It is extensively studied using statistical models or neural models based on the large-scale monolingual source-summary parallel corpus.But there is no cross-lingual parallel corpus, whose source sentence language is different to the summary language, to directly train a cross-lingual ASSUM system.We propose to solve this zero-shot problem by using resource-rich monolingual AS-SUM system to teach zero-shot cross-lingual ASSUM system on both summary word generation and attention.This teaching process is along with a back-translation process which simulates source-summary pairs.Experiments on cross-lingual ASSUM task show that our proposed method is significantly better than pipeline baselines and previous works, and greatly enhances the cross-lingual performances closer to the monolingual performances.We release the code and data at https://github.com/KelleyYin/ Cross-lingual-Summarization.

During the past decades, due to the lack of sufficient labeled data, most studies on crossdomain parsing focus on unsupervised domain adaptation, assuming there is no targetdomain training data.However, unsupervised approaches make limited progress so far due to the intrinsic difficulty of both domain adaptation and parsing.This paper tackles the semi-supervised domain adaptation problem for Chinese dependency parsing, based on two newly-annotated large-scale domain-specific datasets.1 We propose a simple domain embedding approach to merge the sourceand target-domain training data, which is shown to be more effective than both direct corpus concatenation and multi-task learning.In order to utilize unlabeled target-domain data, we employ the recent contextualized word representations and show that a simple fine-tuning procedure can further boost cross-domain parsing accuracy by large margins.

In the literature, existing studies on aspect sentiment classification (ASC) focus on individual non-interactive reviews.This paper extends the research to interactive reviews and proposes a new research task, namely Aspect Sentiment Classification towards Question-Answering (ASC-QA), for real-world applications.This new task aims to predict sentiment polarities for specific aspects from interactive QA style reviews.In particular, a high-quality annotated corpus is constructed for ASC-QA to facilitate corresponding research.On this basis, a Reinforced Bidirectional Attention Network (RBAN) approach is proposed to address two inherent challenges in ASC-QA, i.e., semantic matching between question and answer, and data noise.Experimental results demonstrate the great advantage of the proposed approach to ASC-QA against several state-of-the-art baselines.

The training objective of neural machine translation (NMT) is to minimize the loss between the words in the translated sentences and those in the references. In NMT, there is a natural correspondence between the source sentence and the target sentence. However, this relationship has only been represented using the entire neural network and the training objective is computed in word-level. In this paper, we propose a sentence-level agreement module to directly minimize the difference between the representation of source and target sentence. The proposed agreement module can be integrated into NMT as an additional training objective function and can also be used to enhance the representation of the source sentences. Empirical results on the NIST Chinese-to-English and WMT English-to-German tasks show the proposed agreement module can significantly improve the NMT performance.

Xiangyu Duan, Hongfei Yu, Mingming Yin, Min Zhang, Weihua Luo, Yue Zhang. 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.

Jingjing Wang, Changlong Sun, Shoushan Li, Jiancheng Wang, Luo Si, Min Zhang, Xiaozhong Liu, Guodong Zhou. 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.

Qingrong Xia, Zhenghua Li, Min Zhang. 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.

Xiabing Zhou, Zhongqing Wang, Shoushan Li, Guodong Zhou, Min Zhang. 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.

Jie Zhu, Junhui Li, Muhua Zhu, Longhua Qian, Min Zhang, Guodong Zhou. 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.

The current state-of-the-art dependency parsing approaches employ BiLSTMs to encode input sentences.Motivated by the success of the transformer-based machine translation, this work for the first time applies the self-attention mechanism to dependency parsing as the replacement of the BiLSTM-based encoders, leading to competitive performance on both English and Chinese benchmark data. Based on the detailed error analysis, we then combine the power of both BiLSTM and self-attention via model ensembles, demonstrating their complementary capability of capturing contextual information. Finally, we explore the recently proposed contextualized word representations as extra input features, and further improve the parsing performance.

In Spatial Crowdsourcing (SC) systems, mobile users are enabled to perform spatio-temporal tasks by physically traveling to specified locations with the SC platforms. SC platforms manage the systems and recruit mobile users to contribute to the SC systems, whose commercial success depends on the profit attained from the task requesters. In order to maximize its profit, an SC platform needs an online management mechanism to assign the tasks to suitable workers. How to assign the tasks to workers more cost-effectively with the spatio-temporal constraints is one of the most difficult problems in SC. To deal with this challenge, we propose a novel Profit-driven Task Assignment (PTA) problem, which aims to maximize the profit of the platform. Specifically, we first establish a task reward pricing model with tasks' temporal constraints (i.e., expected completion time and deadline). Then we adopt an optimal algorithm based on tree decomposition to achieve the optimal task assignment and propose greedy algorithms to improve the computational efficiency. Finally, we conduct extensive experiments using real and synthetic datasets, verifying the practicability of our proposed methods.

In human-to-human conversations, the context generally provides several backgrounds and strategic points for the following response. Therefore, many response generation approaches have explored the methodologies to incorporate the context into the encoder-decoder architecture, to generate context-aware responses that are remarkably relevant and cohesive to the given context. However, most approaches pay less attention to semantic interactions implicitly existing within contextual utterances, which are of great importance to capture semantic clues of the given dialog context, indeed. This paper proposes a dynamic working memory mechanism to model long-term semantic hints in the conversation context, by performing semantic interactions between utterances and updating context representation dynamically. Then, the outputs of the dynamic working memory are employed to provide helpful clues for the encoder-decoder architecture to generate responses to the given dialog. We have evaluated the proposed approach on Twitter Customer Service Corpus and OpenSubtitles Corpus, with several automatic evaluation metrics and the human evaluation, and the empirical results show the effectiveness of the proposed method.

Deep Neural Network has achieved great success in many fields. However, many DNN models are both deep and large thereby causing high storage and energy consumption during the training and inference phases. This paper proposes multi-level compression framework. By utilizing cross-layer parameter-reducing techniques ranging from structure compression to weight compression to representation compression, the proposed compression strategy can enable order-of-magnitude reduction in network size for both training and inference with negligible accuracy loss, thereby leading to very high-efficiency and high-accuracy DNN models. Experiments show that the proposed strategy can achieve around 1.8K compression ratio in terms of dense matrices and around 30x for the overall model.

Distant supervised relation extraction is an efficient approach to scale relation extraction to very large corpora, and has been widely used to find novel relational facts from plain text. Recent studies on neural relation extraction have shown great progress on this task via modeling the sentences in low-dimensional spaces, but seldom considered syntax information to model the entities. In this paper, we propose to learn syntax-aware entity embedding for neural relation extraction. First, we encode the context of entities on a dependency tree as sentence-level entity embedding based on tree-GRU. Then, we utilize both intra-sentence and inter-sentence attentions to obtain sentence set-level entity embedding over all sentences containing the focus entity pair. Finally, we combine both sentence embedding and entity embedding for relation classification. We conduct experiments on a widely used real-world dataset and the experimental results show that our model can make full use of all informative instances and achieve state-of-the-art performance of relation extraction.

Neural machine translation (NMT) suffers a performance deficiency when a limited vocabulary fails to cover the source or target side adequately, which happens frequently when dealing with morphologically rich languages. To address this problem, previous work focused on adjusting translation granularity or expanding the vocabulary size. However, morphological information is relatively under-considered in NMT architectures, which may further improve translation quality. We propose a novel method, which can not only reduce data sparsity but also model morphology through a simple but effective mechanism. By predicting the stem and suffix separately during decoding, our system achieves an improvement of up to 1.98 BLEU compared with previous work on English to Russian translation. Our method is orthogonal to different NMT architectures and stably gains improvements on various domains.

To quickly obtain new labeled data, we can choose crowdsourcing as an alternative way at lower cost in a short time. But as an exchange, crowd annotations from non-experts may be of lower quality than those from experts. In this paper, we propose an approach to performing crowd annotation learning for Chinese Named Entity Recognition (NER) to make full use of the noisy sequence labels from multiple annotators. Inspired by adversarial learning, our approach uses a common Bi-LSTM and a private Bi-LSTM for representing annotator-generic and -specific information. The annotator-generic information is the common knowledge for entities easily mastered by the crowd. Finally, we build our Chinese NE tagger based on the LSTM-CRF model. In our experiments, we create two data sets for Chinese NER tasks from two domains. The experimental results show that our system achieves better scores than strong baseline systems.

Treebank conversion is a straightforward and effective way to exploit various heterogeneous treebanks for boosting parsing accuracy.However, previous work mainly focuses on unsupervised treebank conversion and makes little progress due to the lack of manually labeled data where each sentence has two syntactic trees complying with two different guidelines at the same time, referred as bi-tree aligned data.In this work, we for the first time propose the task of supervised treebank conversion.First, we manually construct a bi-tree aligned dataset containing over ten thousand sentences.Then, we propose two simple yet effective treebank conversion approaches (pattern embedding and treeLSTM) based on the state-of-the-art deep biaffine parser.Experimental results show that 1) the two approaches achieve comparable conversion accuracy, and 2) treebank conversion is superior to the widely used multi-task learning framework in multiple treebank exploitation and leads to significantly higher parsing accuracy.* The first two (student) authors make equal contributions to this work.Zhenghua is the correspondence author.

In the popular sequence to sequence (seq2seq) neural machine translation (NMT), there exist many weighted sum models (WSMs), each of which takes a set of input and generates one output. However, the weights in a WSM are independent of each other and fixed for all inputs, suggesting that by ignoring different needs of inputs, the WSM lacks effective control on the influence of each input. In this paper, we propose adaptive weighting for WSMs to control the contribution of each input. Specifically, we apply adaptive weighting for both GRU and the output state in NMT. Experimentation on Chinese-to-English translation and English-to-German translation demonstrates that the proposed adaptive weighting is able to much improve translation accuracy by achieving significant improvement of 1.49 and 0.92 BLEU points for the two translation tasks. Moreover, we discuss in-depth on what type of information is encoded in the encoder and how information influences the generation of target words in the decoder.

Question-Answer (QA) matching is a fundamental task in the Natural Language Processing community. In this paper, we first build a novel QA matching corpus with informal text which is collected from a product reviewing website. Then, we propose a novel QA matching approach, namely One vs. Many Matching, which aims to address the novel scenario where one question sentence often has an answer with multiple sentences. Furthermore, we improve our matching approach by employing both word-level and sentence-level attentions for solving the noisy problem in the informal text. Empirical studies demonstrate the effectiveness of the proposed approach to question-answer matching.

A bottleneck problem with Chinese named entity recognition (NER) in new domains is the lack of annotated data. One solution is to utilize the method of distant supervision, which has been widely used in relation extraction, to automatically populate annotated training data without humancost. The distant supervision assumption here is that if a string in text is included in a predefined dictionary of entities, the string might be an entity. However, this kind of auto-generated data suffers from two main problems: incomplete and noisy annotations, which affect the performance of NER models. In this paper, we propose a novel approach which can partially solve the above problems of distant supervision for NER. In our approach, to handle the incomplete problem, we apply partial annotation learning to reduce the effect of unknown labels of characters. As for noisy annotation, we design an instance selector based on reinforcement learning to distinguish positive sentences from auto-generated annotations. In experiments, we create two datasets for Chinese named entity recognition in two domains with the help of distant supervision. The experimental results show that the proposed approach obtains better performance than the comparison systems on both two datasets.

Chenlin Shen, Changlong Sun, Jingjing Wang, Yangyang Kang, Shoushan Li, Xiaozhong Liu, Luo Si, Min Zhang, Guodong Zhou. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 2018.

We tackle discourse-level relation recognition, a problem of determining semantic relations between text spans.Implicit relation recognition is challenging due to the lack of explicit relational clues.The increasingly popular neural network techniques have been proven effective for semantic encoding, whereby widely employed to boost semantic relation discrimination.However, learning to predict semantic relations at a deep level heavily relies on a great deal of training data, but the scale of the publicly available data in this field is limited.In this paper, we follow Rutherford and Xue (2015) to expand the training data set using the corpus of explicitly-related arguments, by arbitrarily dropping the overtly presented discourse connectives.On the basis, we carry out an experiment of sampling, in which a simple active learning approach is used, so as to take the informative instances for data expansion.The goal is to verify whether the selective use of external data not only reduces the time consumption of retraining but also ensures a better system performance.Using the expanded training data, we retrain a convolutional neural network (CNN) based classifer which is a simplified version of Qin et al. (2016)'s stacking gated relation recognizer.Experimental results show that expanding the training set with small-scale carefully-selected external data yields substantial performance gain, with the improvements of about 4% for accuracy and 3.6% for F-score.This allows a weak classifier to achieve a comparable performance against the state-of-the-art systems.

Although the Transformer translation model (Vaswani et al., 2017) has achieved state-ofthe-art performance in a variety of translation tasks, how to use document-level context to deal with discourse phenomena problematic for Transformer still remains a challenge.In this work, we extend the Transformer model with a new context encoder to represent document-level context, which is then incorporated into the original encoder and decoder.As large-scale document-level parallel corpora are usually not available, we introduce a two-step training method to take full advantage of abundant sentence-level parallel corpora and limited document-level parallel corpora.Experiments on the NIST Chinese-English datasets and the IWSLT French-English datasets show that our approach improves over Transformer significantly. 1

Aspect sentiment classification, a challenging task in sentiment analysis, has been attracting more and more attention in recent years. In this paper, we highlight the need for incorporating the importance degrees of both words and clauses inside a sentence and propose a hierarchical network with both word-level and clause-level attentions to aspect sentiment classification. Specifically, we first adopt sentence-level discourse segmentation to segment a sentence into several clauses. Then, we leverage multiple Bi-directional LSTM layers to encode all clauses and propose a word-level attention layer to capture the importance degrees of words in each clause. Third and finally, we leverage another Bi-directional LSTM layer to encode the outputs from the former layers and propose a clause-level attention layer to capture the importance degrees of all the clauses inside a sentence. Experimental results on the laptop and restaurant datasets from SemEval-2015 demonstrate the effectiveness of our proposed approach to aspect sentiment classification.

Recently, researchers have shown an increasing interest in incorporating linguistic knowledge into neural machine translation (NMT). To this end, previous works choose either to alter the architecture of NMT encoder to incorporate syntactic information into the translation model, or to generalize the embedding layer of the encoder to encode additional linguistic features. The former approach mainly focuses on injecting the syntactic structure of the source sentence into the encoding process, leading to a complicated model that lacks the flexibility to incorporate other types of knowledge. The latter extends word embeddings by considering additional linguistic knowledge as features to enrich the word representation. It thus does not explicitly balance the contribution from word embeddings and the contribution from additional linguistic knowledge. To address these limitations, this paper proposes a knowledge-aware NMT approach that models additional linguistic features in parallel to the word feature. The core idea is that we propose modeling a series of linguistic features at the word level (knowledge block) using a recurrent neural network (RNN). And in sentence level, those word-corresponding feature blocks are further encoded using a RNN encoder. In decoding, we propose a knowledge gate and an attention gate to dynamically control the proportions of information contributing to the generation of target words from different sources. Extensive experiments show that our approach is capable of better accounting for importance of additional linguistic, and we observe significant improvements from 1.0 to 2.3 BLEU points on Chinese$\leftrightarrow$English and English$\rightarrow$German translation tasks.

Automatic chatbots (also known as chat-agents) have attracted much attention from both researching and industrial fields. Generally, the semantic relevance between users' queries and the corresponding responses is considered as the essential element for conversation modeling in both generation and ranking based chat systems. By contrast, it is a nontrivial task to adopt the users' information, such as preference, social role, etc., into conversational models reasonably, while users' profiles play a significant role in the procedure of conversations by providing the implicit contexts. This paper aims to address the personalized response ranking task by incorporating user profiles into the conversation model. In our approach, users' personalized representations are latently learned from the contents posted by them via a two-branch neural network. After that, a deep neural network architecture is further presented to learn the fusion representation of posts, responses, and personal information. In this way, the proposed model could understand conversations from the users' perspective; hence, the more appropriate responses are selected for a specified person. The experimental results on two datasets from social network services demonstrate that our approach is hopeful to represent users' personal information implicitly based on user generated contents, and it is promising to perform as an important component in chatbots to select the personalized responses for each user.

Neural machine translation (NMT) has gained more and more attention in recent years, mainly due to its simplicity yet state-of-the-art performance. However, previous research has shown that NMT suffers from several limitations: source coverage guidance, translation of rare words, and the limited vocabulary, while statistical machine translation (SMT) has complementary properties that correspond well to these limitations. It is straightforward to improve the translation performance by combining the advantages of two kinds of models. This paper proposes a general framework for incorporating the SMT word knowledge into NMT to alleviate above word-level limitations. In our framework, the NMT decoder makes more accurate word prediction by referring to the SMT word recommendations in both training and testing phases. Specifically, the SMT model offers informative word recommendations based on the NMT decoding information. Then, we use the SMT word predictions as prior knowledge to adjust the NMT word generation probability, which unitizes a neural network based classifier to digest the discrete word knowledge. In this paper, we use two model variants to implement the framework, one with a gating mechanism and the other with a direct competition mechanism. Experimental results on Chinese-to-English and English-to-German translation tasks show that the proposed framework can take advantage of the SMT word knowledge and consistently achieve significant improvements over NMT and SMT baseline systems.

In this paper, we propose an analysis mechanism-based structured analysis discriminative dictionary learning analysis discriminative dictionary learning, framework. The ADDL seamlessly integrates ADDL, analysis representation, and analysis classifier training into a unified model. The applied analysis mechanism can make sure that the learned dictionaries, representations, and linear classifiers over different classes are independent and discriminating as much as possible. The dictionary is obtained by minimizing a reconstruction error and an analytical incoherence promoting term that encourages the subdictionaries associated with different classes to be independent. To obtain the representation coefficients, ADDL imposes a sparse -norm constraint on the coding coefficients instead of using or norm, since the - or -norm constraint applied in most existing DL criteria makes the training phase time consuming. The code-extraction projection that bridges data with the sparse codes by extracting special features from the given samples is calculated via minimizing a sparse code approximation term. Then we compute a linear classifier based on the approximated sparse codes by an analysis mechanism to simultaneously consider the classification and representation powers. Thus, the classification approach of our model is very efficient, because it can avoid the extra time-consuming sparse reconstruction process with trained dictionary for each new test data as most existing DL algorithms. Simulations on real image databases demonstrate that our ADDL model can obtain superior performance over other state of the arts.

Neural Machine Translation (NMT) is a new approach to machine translation that has made great progress in recent years. However, recent studies show that NMT generally produces fluent but inadequate translations (Tu et al. 2016b; 2016a; He et al. 2016; Tu et al. 2017). This is in contrast to conventional Statistical Machine Translation (SMT), which usually yields adequate but non-fluent translations. It is natural, therefore, to leverage the advantages of both models for better translations, and in this work we propose to incorporate SMT model into NMT framework. More specifically, at each decoding step, SMT offers additional recommendations of generated words based on the decoding information from NMT (e.g., the generated partial translation and attention history). Then we employ an auxiliary classifier to score the SMT recommendations and a gating function to combine the SMT recommendations with NMT generations, both of which are jointly trained within the NMT architecture in an end-to-end manner. Experimental results on Chinese-English translation show that the proposed approach achieves significant and consistent improvements over state-of-the-art NMT and SMT systems on multiple NIST test sets.

Even though a linguistics-free sequence to sequence model in neural machine translation (NMT) has certain capability of implicitly learning syntactic information of source sentences, this paper shows that source syntax can be explicitly incorporated into NMT effectively to provide further improvements.Specifically, we linearize parse trees of source sentences to obtain structural label sequences.On the basis, we propose three different sorts of encoders to incorporate source syntax into NMT: 1) Parallel RNN encoder that learns word and label annotation vectors parallelly; 2) Hierarchical RNN encoder that learns word and label annotation vectors in a two-level hierarchy; and 3) Mixed RNN encoder that stitchingly learns word and label annotation vectors over sequences where words and labels are mixed.Experimentation on Chinese-to-English translation demonstrates that all the three proposed syntactic encoders are able to improve translation accuracy.It is interesting to note that the simplest RNN encoder, i.e., Mixed RNN encoder yields the best performance with an significant improvement of 1.4 BLEU points.Moreover, an in-depth analysis from several perspectives is provided to reveal how source syntax benefits NMT.

Traditionally, word segmentation (WS) adopts the single-granularity formalism, where a sentence corresponds to a single word sequence.However, Sproat et al. (1996) show that the inter-nativespeaker consistency ratio over Chinese word boundaries is only 76%, indicating single-grained WS (SWS) imposes unnecessary challenges on both manual annotation and statistical modeling.Moreover, WS results of different granularities can be complementary and beneficial for high-level applications.This work proposes and addresses multi-grained WS (MWS).First, we build a large-scale pseudo MWS dataset for model training and tuning by leveraging the annotation heterogeneity of three SWS datasets.Then we manually annotate 1,500 test sentences with true MWS annotations.Finally, we propose three benchmark approaches by casting MWS as constituent parsing and sequence labeling.Experiments and analysis lead to many interesting findings.