Zhaoxuan Tan, Zixuan Zhang, Haoyang Wen, Zheng Li, Rongzhi Zhang, Pei Chen, Fengran Mo, Zheyuan Liu, Qingkai Zeng, Qingyu Yin, Meng Jiang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Abstract Prior works in multi-objective reinforcement learning typically use linear reward scalarization with fixed weights, which provably fail to capture non-convex Pareto fronts and thus yield suboptimal results. This limitation becomes especially critical in online preference alignment for large language models. Here, stochastic trajectories generated by parameterized policies create highly non-linear and non-convex mappings from parameters to objectives that no single static weighting scheme can find optimal trade-offs. We address this limitation by introducing dynamic reward weighting, which adaptively adjusts reward weights during the online reinforcement learning process. Unlike existing approaches that rely on fixed-weight interpolation, our dynamic weighting continuously balances and prioritizes objectives in training, facilitating effective exploration of Pareto fronts in objective space. We introduce two approaches of increasing sophistication and generalizability: hypervolume-guided weight adaptation and gradient-based weight optimization, offering a versatile toolkit for online multi-objective alignment. Our extensive experiments demonstrate their compatibility with commonly used online reinforcement learning algorithms, effectiveness across multiple datasets, and applicability to different model families, consistently achieving Pareto dominant solutions with fewer training steps than fixed-weight linear scalarization baselines.

The safety alignment of large language models (LLMs) remains vulnerable, as their initial behavior can be easily jailbroken by even relatively simple attacks.Since infilling a fixed template between the input instruction and initial model output is a common practice for existing LLMs, we hypothesize that this template is a key factor behind their vulnerabilities: LLMs' safety-related decision-making overly relies on the aggregated information from the template region, which largely influences these models' safety behavior.We refer to this issue as template-anchored safety alignment.In this paper, we conduct extensive experiments and verify that template-anchored safety alignment is widespread across various aligned LLMs.Our mechanistic analyses demonstrate how it leads to models' susceptibility when encountering inference-time jailbreak attacks.Furthermore, we show that detaching safety mechanisms from the template region is promising in mitigating vulnerabilities to jailbreak attacks.We encourage future research to develop more robust safety alignment techniques that reduce reliance on the template region.

Currently, most multimodal studies are based on large language models (LLMs) with quadratic-complexity Transformer architectures.While linear models like RNNs enjoy low inference costs, their application has been largely limited to the text-only modality.This work explores the capabilities of modern RNN architectures in multimodal contexts.We propose ModRWKV-a decoupled multimodal framework built upon the RWKV7 architecture as its LLM backbone-which achieves multisource information fusion through dynamically adaptable heterogeneous modality encoders.We designed the multimodal modules in Mod-RWKV with an extremely lightweight architecture and, through extensive experiments, identified a configuration that achieves an optimal balance between performance and computational efficiency.ModRWKV leverages the pretrained weights of the RWKV7 LLM for initialization, which significantly accelerates multimodal training.Comparative experiments with different pretrained checkpoints further demonstrate that such initialization plays a crucial role in enhancing the model's ability to understand multimodal signals.Supported by extensive experiments, we conclude that modern RNN architectures present a viable alternative to Transformers in the domain of multimodal large language models (MLLMs).Furthermore, we identify the optimal configuration of the ModRWKV architecture through systematic exploration.

Large language models (LLMs) have achieved remarkable performance on vari- ous natural language tasks. However, they are trained on static corpora and their knowledge can become outdated quickly in the fast-changing world. This moti- vates the development of knowledge editing methods designed to update certain knowledge in LLMs without changing unrelated others. To make selective edits, previous efforts often sought to update a small amount of parameters in some spe- cific layer(s) of a LLM. Nonetheless, in challenging scenarios, they still fall short in making successful edits while preserving knowledge irrelevant to the updates simultaneously, resulting in a notable editing-locality trade-off. In this work, we question if the trade-offs are caused by the fact that parameter-based updates have a global effect, i.e., edited parameters affect all inputs indiscriminately. In light of this, we explore the feasibility of representation fine-tuning, which applied some linear update to a few representations in a learned subspace, for knowledge edit- ing. While being effective to enhance an LLM’s general ability as demonstrated in the previous work, we theoretically show that this linear update imposes a tension in editing-locality trade-off. Subsequently, BaFT is proposed to break the linear- ity. BaFT computes a weight for each basis that spans a dimension of the subspace based on the input representation. This input-dependent weighting mechanism al- lows BaFT to manage different types of knowledge in an adaptive way, thereby achieving a better editing-locality trade-off. Experiments on three LLMs with five editing benchmarks in diverse scenarios show the superiority of our method.

Large language models (LLMs) have achieved remarkable performance on various natural language tasks. However, they are trained on static corpora and their knowledge can become outdated quickly in the fast-changing world. This motivates the development of knowledge editing (KE) to update specific knowledge in LLMs without changing unrelated others or compromising their pre-trained capabilities. Previous efforts sought to update a small amount of parameters of a LLM and proved effective for making selective updates. Nonetheless, the edited LLM often exhibits degraded ability to reason about the new knowledge. In this work, we identify a key issue: heterogeneous token overfitting (HTO), where the LLM overfits different tokens in the provided knowledge at varying rates. To tackle this, we propose OVERTONE, a token-level smoothing method that mitigates HTO by adaptively refining the target distribution. Theoretically, OVERTONE offers better parameter updates with negligible computation overhead. It also induces an implicit DPO but does not require preference data pairs. Extensive experiments across four editing methods, two LLMs, and diverse scenarios demonstrate the effectiveness and versatility of our method.

The alignment of large language models (LLMs) with human preferences remains a key challenge. While post-training techniques like Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) have achieved notable success, they often experience computational inefficiencies and training instability. In this paper, we propose Feature-level constrained Preference Optimization (FPO), a novel method designed to simplify the alignment process while ensuring stability. FPO leverages pre-trained Sparse Autoencoders (SAEs) and introduces feature-level constraints, allowing for efficient, sparsity-enforced alignment. Our approach enjoys efficiency by using sparse features activated in a well-trained sparse autoencoder and the quality of sequential KL divergence by using the feature-level offline reference. Experimental results on benchmark datasets demonstrate that FPO achieves an above 5% absolute improvement in win rate with much lower computational cost compared to state-of-the-art baselines, making it a promising solution for efficient and controllable LLM alignments.

Zhihan Zhang, Shiyang Li, Zixuan Zhang, Xin Liu, Haoming Jiang, Xianfeng Tang, Yifan Gao, Zheng Li, Haodong Wang, Zhaoxuan Tan, Yichuan Li, Qingyu Yin, Bing Yin, Meng Jiang. 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.

Recent language models exhibit strong reasoning capabilities, yet the influence of long-context capacity on reasoning remains underexplored. In this work, we hypothesize that current limitations in reasoning stem, in part, from insufficient long-context capacity, motivated by empirical observations such as i) higher context window length often leads to stronger reasoning performance, and ii) failed reasoning cases resemble failed long-context cases. To test this hypothesis, we examine whether enhancing a model’s long-context ability before Supervised Fine-Tuning (SFT) leads to improved reasoning performance. Specifically, we compared models with identical architectures and fine-tuning data but varying levels of long-context capacity. Our results reveal a consistent trend: models with stronger long-context capacity achieve significantly higher accuracy on reasoning benchmarks after SFT. Notably, these gains persist even on tasks with short input lengths, indicating that long-context training offers generalizable benefits for reasoning performance. These findings suggest that long-context modeling is not just essential for processing lengthy inputs, but also serves as a critical foundation for reasoning. We advocate for treating long-context capacity as a first-class objective in the design of future language models.

Fenglin Liu, Zheng Li, Hongjian Zhou, Qingyu Yin, Jingfeng Yang, Xianfeng Tang, Chen Luo, Ming Zeng, Haoming Jiang, Yifan Gao, Priyanka Nigam, Sreyashi Nag, Bing Yin, Yining Hua, Xuan Zhou, Omid Rohanian, Anshul Thakur, Lei Clifton, David A. Clifton. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Baixuan Xu, Weiqi Wang, Haochen Shi, Wenxuan Ding, Huihao Jing, Tianqing Fang, Jiaxin Bai, Xin Liu, Changlong Yu, Zheng Li, Chen Luo, Qingyu Yin, Bing Yin, Long Chen, Yangqiu Song. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Developing a universal model that can effectively harness heterogeneous resources and respond to a wide range of personalized needs has been a longstanding community aspiration. Our daily choices, especially in domains like fashion and retail, are substantially shaped by multi-modal data, such as pictures and textual descriptions. These modalities not only offer intuitive guidance but also cater to personalized user preferences. However, the predominant personalization approaches mainly focus on ID or text-based recommendation problems, failing to comprehend the information spanning various tasks or modalities. In this paper, our goal is to establish a Unified paradigm for Multi-modal Personalization systems (UniMP), which effectively leverages multi-modal data while eliminating the complexities associated with task- and modality-specific customization. We argue that the advancements in foundational generative modeling have provided the flexibility and effectiveness necessary to achieve the objective. In light of this, we develop a generic and extensible personalization generative framework, that can handle a wide range of personalized needs including item recommendation, product search, preference prediction, explanation generation, and further user-guided image generation. Our methodology enhances the capabilities of foundational language models for personalized tasks by seamlessly ingesting interleaved cross-modal user history information, ensuring a more precise and customized experience for users. To train and evaluate the proposed multi-modal personalized tasks, we also introduce a novel and comprehensive benchmark covering a variety of user requirements. Our experiments on the real-world benchmark showcase the model's potential, outperforming competitive methods specialized for each task.

Existing Large Language Models (LLMs) usually remain static after deployment, which might make it hard to inject new knowledge into the model. We aim to build models containing a considerable portion of self-updatable parameters, enabling the model to integrate new knowledge effectively and efficiently. To this end, we introduce MEMORYLLM, a model that comprises a transformer and a fixed-size memory pool within the latent space of the transformer. MEMORYLLM can self-update with text knowledge and memorize the knowledge injected earlier. Our evaluations demonstrate the ability of MEMORYLLM to effectively incorporate new knowledge, as evidenced by its performance on model editing benchmarks. Meanwhile, the model exhibits long-term information retention capacity, which is validated through our custom-designed evaluations and long-context benchmarks. MEMORYLLM also shows operational integrity without any sign of performance degradation even after nearly a million memory updates. Our code and model are open-sourced at https://github.com/wangyu-ustc/MemoryLLM.

The decoder-only Transformer architecture with causal masking and relative position encoding (RPE) has become the de facto choice in language modeling. Despite its exceptional performance across various tasks, we have identified two limitations: First, it prevents all attended tokens from having zero weights during the softmax stage, even if the current embedding has sufficient self-contained information. This compels the model to assign disproportional excessive attention to specific tokens. Second, RPE-based Transformers are not universal approximators due to their limited capacity at encoding absolute positional information, which limits their application in position-critical tasks. In this work, we propose StableMask: a parameter-free method to address both limitations by refining the causal mask. It introduces pseudo-attention values to balance attention distributions and encodes absolute positional information via a progressively decreasing mask ratio. StableMask's effectiveness is validated both theoretically and empirically, showing significant enhancements in language models with parameter sizes ranging from 71M to 1.4B across diverse datasets and encoding methods. We further show that it supports integration with existing optimization techniques, making it easily usable in practical applications.

Understanding user intentions is essential for improving product recommendations, navigation suggestions, and query reformulations. However, user intentions can be intricate, involving multiple sessions and attribute requirements connected by logical operators such as And, Or, and Not. For instance, a user may search for Nike or Adidas running shoes across various sessions, with a preference for purple. In another example, a user may have purchased a mattress in a previous session and is now looking for a matching bed frame without intending to buy another mattress. Existing research on session understanding has not adequately addressed making product or attribute recommendations for such complex intentions. In this paper, we present the task of logical session complex query answering (LS-CQA), where sessions are treated as hyperedges of items, and we frame the problem of complex intention understanding as an LS-CQA task on an aggregated hypergraph of sessions, items, and attributes. This is a unique complex query answering task with sessions as ordered hyperedges. We also introduce a new model, the Logical Session Graph Transformer (LSGT), which captures interactions among items across different sessions and their logical connections using a transformer structure. We analyze the expressiveness of LSGT and prove the permutation invariance of the inputs for the logical operators. By evaluating LSGT on three datasets, we demonstrate that it achieves state-of-the-art results.

Xiusi Chen, Hongzhi Wen, Sreyashi Nag, Chen Luo, Qingyu Yin, Ruirui Li, Zheng Li, Wei Wang. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Online shopping is a complex multi-task, few-shot learning problem with a wide and evolving range of entities, relations, and tasks. However, existing models and benchmarks are commonly tailored to specific tasks, falling short of capturing the full complexity of online shopping. Large Language Models (LLMs), with their multi-task and few-shot learning abilities, have the potential to profoundly transform online shopping by alleviating task-specific engineering efforts and by providing users with interactive conversations. Despite the potential, LLMs face unique challenges in online shopping, such as domain-specific concepts, implicit knowledge, and heterogeneous user behaviors. Motivated by the potential and challenges, we propose Shopping MMLU, a diverse multi-task online shopping benchmark derived from real-world Amazon data. Shopping MMLU consists of 57 tasks covering 4 major shopping skills: concept understanding, knowledge reasoning, user behavior alignment, and multi-linguality, and can thus comprehensively evaluate the abilities of LLMs as general shop assistants. With Shoppping MMLU, we benchmark over 20 existing LLMs and uncover valuable insights about practices and prospects of building versatile LLM-based shop assistants. Shopping MMLU can be publicly accessed at https://github.com/KL4805/ShoppingMMLU. In addition, with Shopping MMLU, we are hosting a competition in KDD Cup 2024 with over 500 participating teams. The winning solutions and the associated workshop can be accessed at our website https://amazon-kddcup24.github.io/.

Textbook question answering (TQA) is the task of correctly answering diagram or nondiagram (ND) questions given large multimodal contexts consisting of abundant essays and diagrams. In real-world scenarios, an explainable TQA system plays a key role in deepening humans' understanding of learned knowledge. However, there is no work to investigate how to provide explanations currently. To address this issue, we devise a novel architecture toward span-level eXplanations for TQA (XTQA). In this article, spans are the combinations of sentences within a paragraph. The key idea is to consider the entire textual context of a lesson as candidate evidence and then use our proposed coarse-to-fine grained explanation extracting (EE) algorithm to narrow down the evidence scope and extract the span-level explanations with varying lengths for answering different questions. The EE algorithm can also be integrated into other TQA methods to make them explainable and improve the TQA performance. Experimental results show that XTQA obtains the best overall explanation result [mean intersection over union (mIoU)] of 52.38% on the first 300 questions of CK12-QA test splits, demonstrating the explainability of our method (ND: 150 and diagram: 150). The results also show that XTQA achieves the best TQA performance of 36.46% and 36.95% on the aforementioned splits. We have released our code in https://github.com/dr-majie/opentqa.

Aiming at a better understanding of the search goals in the user search sessions, recent query recommender systems explicitly model the reformulations of queries, which hopes to estimate the intents behind these reformulations and thus benefit the next-query recommendation. However, in real-world e-commercial search scenarios, user intents are much more complicated and may evolve dynamically. Existing methods merely consider trivial reformulation intents from semantic aspects and fail to model dynamic reformulation intent flows in search sessions, leading to sub-optimal capacities to recommend desired queries. To deal with these limitations, we first explicitly define six types of query reformulation intents according to the desired products of two consecutive queries. We then apply two self-attentive encoders on top of two pre-trained large language models to learn the transition dynamics from semantic query and intent reformulation sequences, respectively. We develop an intent-aware query decoder to utilize the predicted intents for suggesting the next queries. We instantiate such a framework with an Intent-aware Variational AutoEncoder (IVAE) under deployment at Amazon. We conduct comprehensive experiments on two real-world e-commercial datasets from Amazon and one public dataset from BestBuy. Specifically, IVAE improves the Recall@15 by 25.44% and 60.47% on two Amazon datasets and 13.91% on BestBuy, respectively.

A complex logic query in a knowledge graph refers to a query expressed in logic form that conveys a complex meaning, such as where did the Canadian Turing award winner graduate from? Knowledge graph reasoning-based applications, such as dialogue systems and interactive search engines, rely on the ability to answer complex logic queries as a fundamental task. In most knowledge graphs, edges are typically used to either describe the relationships between entities or their associated attribute values. An attribute value can be in categorical or numerical format, such as dates, years, sizes, etc. However, existing complex query answering (CQA) methods simply treat numerical values in the same way as they treat entities. This can lead to difficulties in answering certain queries, such as which Australian Pulitzer award winner is born before 1927, and which drug is a pain reliever and has fewer side effects than Paracetamol. In this work, inspired by the recent advances in numerical encoding and knowledge graph reasoning, we propose numerical complex query answering. In this task, we introduce new numerical variables and operations to describe queries involving numerical attribute values. To address the difference between entities and numerical values, we also propose the framework of Number Reasoning Network (NRN) for alternatively encoding entities and numerical values into separate encoding structures. During the numerical encoding process, NRN employs a parameterized density function to encode the distribution of numerical values. During the entity encoding process, NRN uses established query encoding methods for the original CQA problem. Experimental results show that NRN consistently improves various query encoding methods on three different knowledge graphs and achieves state-of-the-art results.

Rui Feng, Chen Luo, Qingyu Yin, Bing Yin, Tuo Zhao, Chao Zhang. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

With the continuous development of multi-access edge computing (MEC) and artificial intelligence (AI), a new paradigm for 6G, namely endogenous intelligence (EI)-based MEC is created. Hence, a novel EI-based MEC scheme is proposed in this paper. Firstly, we present an EI-based MEC architecture, which deepens EI into MEC architecture. Secondly, referring to service-based architecture (SBA) for 5G core network (5GC), we decouple the edge intelligent services into multiple network functions (NFs) to enhance the flexibility of EI-based MEC architecture. Thirdly, for specific service types, we design MEC templates and instantiation schemes to complete the reconfiguration of EI-based MEC. Finally, we establish a testbed, and experimental results demonstrate that our proposed EI-based MEC can provide users with more reliable and agile edge intelligent services and effectively improve the Quality of Service (QoS).

Query understanding models extract attributes from search queries, like color, product type, brand, etc. Search engines rely on these attributes for ranking, advertising, and recommendation, etc. However, product search queries are usually short, three or four words on average. This information shortage limits the search engine’s power to provide high-quality services.

With the increasing demands on e-commerce platforms, numerous user action history is emerging. Those enriched action records are vital to understand users’ interests and intents. Recently, prior works for user behavior prediction mainly focus on the interactions with product-side information. However, the interactions with search queries, which usually act as a bridge between users and products, are still under investigated. In this paper, we explore a new problem named temporal event forecasting, a generalized user behavior prediction task in a unified query product evolutionary graph, to embrace both query and product recommendation in a temporal manner. To fulfill this setting, there involves two challenges: (1) the action data for most users is scarce; (2) user preferences are dynamically evolving and shifting over time. To tackle those issues, we propose a novel Retrieval-Enhanced Temporal Event (RETE) forecasting framework. Unlike existing methods that enhance user representations via roughly absorbing information from connected entities in the whole graph, RETE efficiently and dynamically retrieves relevant entities centrally on each user as high-quality subgraphs, preventing the noise propagation from the densely evolutionary graph structures that incorporate abundant search queries. And meanwhile, RETE autoregressively accumulates retrieval-enhanced user representations from each time step, to capture evolutionary patterns for joint query and product prediction. Empirically, extensive experiments on both the public benchmark and four real-world industrial datasets demonstrate the effectiveness of the proposed RETE method.

Table-based fact verification task aims to verify whether the given statement is supported by the given semi-structured table. Symbolic reasoning with logical operations plays a crucial role in this task. Existing methods leverage programs that contain rich logical information to enhance the verification process. However, due to the lack of fully supervised signals in the program generation process, spurious programs can be derived and employed, which leads to the inability of the model to catch helpful logical operations. To address the aforementioned problems, in this work, we formulate the table-based fact verification task as an evidence retrieval and reasoning framework, proposing the Logic-level Evidence Retrieval and Graph-based Verification network (LERGV). Specifically, we first retrieve logic-level program-like evidence from the given table and statement as supplementary evidence for the table. After that, we construct a logic-level graph to capture the logical relations between entities and functions in the retrieved evidence, and design a graph-based verification network to perform logic-level graph-based reasoning based on the constructed graph to classify the final entailment relation. Experimental results on the large-scale benchmark TABFACT show the effectiveness of the proposed approach.

Table-based fact verification is expected to perform both linguistic reasoning and symbolic reasoning.Existing methods lack attention to take advantage of the combination of linguistic information and symbolic information.In this work, we propose HeterTFV, a graph-based reasoning approach, that learns to combine linguistic information and symbolic information effectively.We first construct a program graph to encode programs, a kind of LISP-like logical form, to learn the semantic compositionality of the programs.Then we construct a heterogeneous graph to incorporate both linguistic information and symbolic information by introducing program nodes into the heterogeneous graph.Finally, we propose a graph-based reasoning approach to reason over the multiple types of nodes to make an effective combination of both types of information.Experimental results on a large-scale benchmark dataset TABFACT illustrate the effect of our approach.

Semantic information that has been proven to be necessary to the resolution of common noun phrases is typically ignored by most existing Chinese zero pronoun resolvers. This is because that zero pronouns convey no descriptive information, which makes it almost impossible to calculate semantic similarities between the zero pronoun and its candidate antecedents. Moreover, most of traditional approaches are based on the single-candidate model, which considers the candidate antecedents of a zero pronoun in isolation and thus overlooks their reciprocities. To address these problems, we first propose a neural-network-based zero pronoun resolver ( NZR ) that is capable of generating vector-space semantics of zero pronouns and candidate antecedents. On the basis of NZR , we develop the collaborative filtering-based framework for Chinese zero pronoun resolution task, exploring the reciprocities between the candidate antecedents of a zero pronoun to more rationally re-estimate their importance. Experimental results on the Chinese portion of the OntoNotes 5.0 corpus are encouraging: Our proposed model substantially surpasses the Chinese zero pronoun resolution baseline systems.

Building explainable systems is a critical problem in the field of Natural Language Processing (NLP), since most machine learning models provide no explanations for the predictions. Existing approaches for explainable machine learning systems tend to focus on interpreting the outputs or the connections between inputs and outputs. However, the fine-grained information (e.g. textual explanations for the labels) is often ignored, and the systems do not explicitly generate the human-readable explanations. To solve this problem, we propose a novel generative explanation framework that learns to make classification decisions and generate fine-grained explanations at the same time. More specifically, we introduce the explainable factor and the minimum risk training approach that learn to generate more reasonable explanations. We construct two new datasets that contain summaries, rating scores, and fine-grained reasons. We conduct experiments on both datasets, comparing with several strong neural network baseline systems. Experimental results show that our method surpasses all baselines on both datasets, and is able to generate concise explanations at the same time.

Deep neural network models for Chinese zero pronoun resolution learn semantic information for zero pronoun and candidate antecedents, but tend to be short-sightedthey often make local decisions.They typically predict coreference chains between the zero pronoun and one single candidate antecedent one link at a time, while overlooking their long-term influence on future decisions.Ideally, modeling useful information of preceding potential antecedents is critical when later predicting zero pronoun-candidate antecedent pairs.In this study, we show how to integrate local and global decision-making by exploiting deep reinforcement learning models.With the help of the reinforcement learning agent, our model learns the policy of selecting antecedents in a sequential manner, where useful information provided by earlier predicted antecedents could be utilized for making later coreference decisions.Experimental results on OntoNotes 5.0 dataset show that our technique surpasses the state-of-the-art models.

Recent neural network methods for zero pronoun resolution explore multiple models for generating representation vectors for zero pronouns and their candidate antecedents. Typically, contextual information is utilized to encode the zero pronouns since they are simply gaps that contain no actual content. To better utilize contexts of the zero pronouns, we here introduce the self-attention mechanism for encoding zero pronouns. With the help of the multiple hops of attention, our model is able to focus on some informative parts of the associated texts and therefore produces an efficient way of encoding the zero pronouns. In addition, an attention-based recurrent neural network is proposed for encoding candidate antecedents by their contents. Experiment results are encouraging: our proposed attention-based model gains the best performance on the Chinese portion of the OntoNotes corpus, substantially surpasses existing Chinese zero pronoun resolution baseline systems.

Most existing approaches for zero pronoun resolution are heavily relying on annotated data, which is often released by shared task organizers.Therefore, the lack of annotated data becomes a major obstacle in the progress of zero pronoun resolution task.Also, it is expensive to spend manpower on labeling the data for better performance.To alleviate the problem above, in this paper, we propose a simple but novel approach to automatically generate large-scale pseudo training data for zero pronoun resolution.Furthermore, we successfully transfer the cloze-style reading comprehension neural network model into zero pronoun resolution task and propose a two-step training mechanism to overcome the gap between the pseudo training data and the real one.Experimental results show that the proposed approach significantly outperforms the state-of-the-art systems with an absolute improvements of 3.1% F-score on OntoNotes 5.0 data.

Existing approaches for Chinese zero pronoun resolution typically utilize only syntactical and lexical features while ignoring semantic information.The fundamental reason is that zero pronouns have no descriptive information, which brings difficulty in explicitly capturing their semantic similarities with antecedents.Meanwhile, representing zero pronouns is challenging since they are merely gaps that convey no actual content.In this paper, we address this issue by building a deep memory network that is capable of encoding zero pronouns into vector representations with information obtained from their contexts and potential antecedents.Consequently, our resolver takes advantage of semantic information by using these continuous distributed representations.Experiments on the OntoNotes 5.0 dataset show that the proposed memory network could substantially outperform the state-of-the-art systems in various experimental settings.

Existing approaches for Chinese zero pronoun resolution overlook semantic information. This is because zero pronouns have no descriptive information, which results in difficulty in explicitly capturing their semantic similarities with antecedents. Moreover, when dealing with candidate antecedents, traditional systems simply take advantage of the local information of a single candidate antecedent while failing to consider the underlying information provided by the other candidates from a global perspective. To address these weaknesses, we propose a novel zero pronoun-specific neural network, which is capable of representing zero pronouns by utilizing the contextual information at the semantic level. In addition, when dealing with candidate antecedents, a two-level candidate encoder is employed to explicitly capture both the local and global information of candidate antecedents. We conduct experiments on the Chinese portion of the OntoNotes 5.0 corpus. Experimental results show that our approach substantially outperforms the state-of-the-art method in various experimental settings.