COLING 2025 论文列表

In recent years, fine-grained sentiment analysis in finance has gained significant attention, but the scarcity of entity-level datasets remains a key challenge. To address this, we have constructed the largest English and Chinese financial entity-level sentiment analysis datasets to date. Building on this foundation, we propose a novel two-stage sentiment analysis approach called Self-aware In-context Learning Correction (SILC). The first stage involves fine-tuning a base large language model to generate pseudo-labeled data specific to our task. In the second stage, we train a correction model using a GNN-based example retriever, which is informed by the pseudo-labeled data. This two-stage strategy has allowed us to achieve state-of-the-art performance on the newly constructed datasets, advancing the field of financial sentiment analysis. In a case study, we demonstrate the enhanced practical utility of our data and methods in monitoring the cryptocurrency market. Our datasets and code are available at https://github.com/NLP-Bin/SILC-EFSA.

Multimodal Entity Alignment (MMEA) aims to identify equivalent entities across different multimodal knowledge graphs (MMKGs) by integrating structural information, entity attributes, and visual data, thereby promoting knowledge sharing and deep multimodal data integration. However, existing methods often overlook the deeper connections between multimodal data. They primarily focus on the interactions between neighboring entities in the structural modality while neglecting the interactions between entities in the visual and attribute modalities. To address this, we propose a structure-guided multimodal entity alignment method (SGMEA), which prioritizes structural information from knowledge graphs to enhance the visual and attribute modalities. By fusing multimodal representations, SGMEA improves the accuracy of entity alignment. Experimental results demonstrate that SGMEA achieves stateof-the-art performance across multiple datasets, validating its effectiveness and superiority in practical applications.

Large Language Models (LLMs) demonstrate remarkable emergent abilities across various tasks, yet fall short of complex reasoning and planning tasks. The tree-search-based reasoning methods address this by encouraging the exploration of intermediate steps, surpassing the capabilities of chain-of-thought prompting. However, significant inference latency is introduced due to the systematic exploration and evaluation of multiple thought paths. This paper introduces SEED, a novel and efficient inference framework to improve both runtime speed and GPU memory management concurrently. Based on a scheduled speculative execution, SEED efficiently handles multiple iterations for thought generation and state evaluation, leveraging a rounds-scheduled strategy to manage draft model dispatching. Extensive experimental evaluations on three reasoning datasets demonstrate the superior speedup performance of SEED.

Recent advances in language models (LMs) has driven progress in information retrieval (IR), effectively extracting semantically relevant information. However, they face challenges in balancing computational costs with deeper query-document interactions. To tackle this, we present two mechanisms: 1) a light and effective multi-vector retrieval with sequence compression vectors, dubbed SCV and 2) coarse-to-fine vector search. The strengths of SCV stems from its application of span compressive vectors for scoring. By employing a non-linear operation to examine every token in the document, we abstract these into a span-level representation. These vectors effectively reduce the document’s dimensional representation, enabling the model to engage comprehensively with tokens across the entire collection of documents, rather than the subset retrieved by Approximate Nearest Neighbor. Therefore, our framework performs a coarse single vector search during the inference stage and conducts a fine-grained multi-vector search end-to-end. This approach effectively reduces the cost required for search. We empirically show that SCV achieves the fastest latency compared to other state-of-the-art models and can obtain competitive performance on both in-domain and out-of-domain benchmark datasets.

The rampant spread of cyberbullying content poses a growing threat to societal well-being. However, research on cyberbullying detection in Chinese remains underdeveloped, primarily due to the lack of comprehensive and reliable datasets. Notably, no existing Chinese dataset is specifically tailored for cyberbullying detection. Moreover, while comments play a crucial role within sessions, current session-based datasets often lack detailed, fine-grained annotations at the comment level. To address these limitations, we present a novel Chinese cyberbullying dataset, termed SCCD, which consists of 677 session-level samples sourced from a major social media platform Weibo. Moreover, each comment within the sessions is annotated with fine-grained labels rather than conventional binary class labels. Empirically, we evaluate the performance of various baseline methods on SCCD, highlighting the challenges for effective Chinese cyberbullying detection.

The development of unbiased large language models is widely recognized as crucial, yet existing benchmarks fall short in detecting biases due to limited scope, contamination, and lack of a fairness baseline. SAGED(bias) is the first holistic benchmarking pipeline to address these problems. The pipeline encompasses five core stages: scraping materials, assembling benchmarks, generating responses, extracting numeric features, and diagnosing with disparity metrics. SAGED includes metrics for max disparity, such as impact ratio, and bias concentration, such as Max Z-scores. Noticing that metric tool bias and contextual bias in prompts can distort evaluation, SAGED implements counterfactual branching and baseline calibration for mitigation. For demonstration, we use SAGED on G20 Countries with popular 8b-level models including Gemma2, Llama3.1, Mistral, and Qwen2. With sentiment analysis, we find that while Mistral and Qwen2 show lower max disparity and higher bias concentration than Gemma2 and Llama3.1, all models are notably biased against countries like Russia and (except for Qwen2) China. With further experiments to have models role-playing U.S. presidents, we see bias amplifies and shifts in heterogeneous directions. Moreover, we see Qwen2 and Mistral not engage in role-playing, while Llama3.1 and Gemma2 role-play Trump notably more intensively than Biden and Harris, indicating role-playing performance bias in these models.

Moment Retrieval aims to locate specific video segments related to the given text. Recently, DETR-based methods, originating from Object Detection, have emerged as effective solutions for Moment Retrieval. These approaches focus on multimodal feature fusion and refining Queries composed of span anchor and content embedding. Despite the success, they often overlook the video-text instance related information in Query Initialization and the crucial guidance role of span anchors in Query Refinement, leading to inaccurate predictions. To address this, we propose a novel Span Aware DEtection TRansformer (SA-DETR) that leverages the importance of instance related span anchors. To fully leverage the instance related information, we generate span anchors based on video-text pair rather than using learnable parameters, as is common in conventional DETR-based methods, and supervise them with GT labels. To effectively exploit the correspondence between span anchors and video clips, we enhance content embedding guided by textual features and generate Gaussian mask to modulate the interaction between content embedding and fusion features. Furthermore, we explore the feature alignment across various stages and granularities and apply denoise learning to boost the span awareness of the model. Extensive experiments on QVHighlights, Charades-STA, and TACoS demonstrate the effectiveness of our approach.

Traditional methods for detecting rumors on social media primarily focus on analyzing textual content, often struggling to capture the complexity of online interactions. Recent research has shifted towards leveraging graph neural networks to model the hierarchical conversation structure that emerges during rumor propagation. However, these methods tend to overlook the temporal aspect of rumor propagation and may disregard potential noise within the propagation structure. In this paper, we propose a novel approach that incorporates temporal information by constructing a weighted propagation tree, where the weight of each edge represents the time interval between connected posts. Drawing upon the theory of structural entropy, we transform this tree into a coding tree. This transformation aims to preserve the essential structure of rumor propagation while reducing noise. Finally, we introduce a recursive neural network to learn from the coding tree for rumor veracity prediction. Experimental results on two common datasets demonstrate the superiority of our approach.

Knowledge base question answering (KBQA) is recognized as a challenging task, especially when parsing complex questions into executable logical forms. Traditional semantic parsing (SP)-based approaches exhibit inconsistent performance in handling various complex questions. As large language models (LLMs) have exhibited exceptional reasoning ability and language comprehension, recent studies have employed LLMs for semantic parsing to directly generate logical forms that can be executed on knowledge bases (KBs) to achieve the desired results. However, these methods of relying exclusively on LLMs to ensure grammaticality, faithfulness, and controllability may diminish their effectiveness due to hallucinations in the reasoning process. In this paper, we introduce Rule-KBQA, a framework that employs learned rules to guide the generation of logical forms. The proposed method contains two phases, an induction phase and a deduction phase. In the induction phase, we initially extract rules from the existing data and then employ the Rule-Following Fine-Tuned (RFFT) LLM to generate additional rules, ultimately constructing a comprehensive rule library. In the deduction phase, a symbolic agent, guided by learned rules, explores the environment KB to incrementally construct executable logical forms. Meanwhile, we leverage the discriminative capability of LLMs to evaluate the plausibility of candidate decisions. Extensive experiments indicate that our method achieves competitive results on standard KBQA datasets, clearly demonstrating its effectiveness.

Role-playing systems powered by large language models (LLMs) have become increasingly influential in emotional communication applications. However, these systems are susceptible to character hallucinations, where the model deviates from predefined character roles and generates responses that are inconsistent with the intended persona. This paper presents the first systematic analysis of character hallucination from an attack perspective, introducing the RoleBreak framework. Our framework identifies two core mechanisms—query sparsity and role-query conflict—as key factors driving character hallucination. Leveraging these insights, we construct a novel dataset, RoleBreakEval, to evaluate existing hallucination mitigation techniques. Our experiments reveal that even enhanced models trained to minimize hallucination remain vulnerable to attacks. To address these vulnerabilities, we propose a novel defence strategy, the Narrator Mode, which generates supplemental context through narration to mitigate role-query conflicts and improve query generalization. Experimental results demonstrate that Narrator Mode significantly outperforms traditional refusal-based strategies by reducing hallucinations, enhancing fidelity to character roles and queries, and improving overall narrative coherence.

To ensure reliable performance of Question Answering (QA) systems, evaluation of robustness is crucial. Common evaluation benchmarks commonly only include performance metrics, such as Exact Match (EM) and the F1 score. However, these benchmarks overlook critical factors for the deployment of QA systems. This oversight can result in systems vulnerable to minor perturbations in the input such as typographical errors. While several methods have been proposed to test the robustness of QA models, there has been minimal exploration of these approaches for languages other than English. This study focuses on the robustness evaluation of German language QA models, extending methodologies previously applied primarily to English. The objective is to nurture the development of robust models by defining an evaluation method specifically tailored to the German language. We assess the applicability of perturbations used in English QA models for German and perform a comprehensive experimental evaluation with eight models. The results show that all models are vulnerable to character-level perturbations. Additionally, the comparison of monolingual and multilingual models suggest that the former are less affected by character and word-level perturbations.

Using supervised automatic summarisation methods requires sufficient corpora that include pairs of documents and their summaries. Similarly to many tasks in natural language processing, most of the datasets available for summarization are in English, posing challenges for developing summarization models in other languages. Thus, in this work, we introduce RoLargeSum, a novel large-scale summarization dataset for the Romanian language crawled from various publicly available news websites from Romania and the Republic of Moldova that were thoroughly cleaned to ensure a high-quality standard. RoLargeSum contains more than 615K news articles, together with their summaries, as well as their headlines, keywords, dialect, and other metadata that we found on the targeted websites. We further evaluated the performance of several BART variants and open-source large language models on RoLargeSum for benchmarking purposes. We manually evaluated the results of the best-performing system to gain insight into the potential pitfalls of this data set and future development.

Randomized Controlled Trials (RCTs) are rigorous clinical studies crucial for reliable decision-making, but their credibility can be compromised by bias. The Cochrane Risk of Bias tool (RoB 2) assesses this risk, yet manual assessments are time-consuming and labor-intensive. Previous approaches have employed Large Language Models (LLMs) to automate this process. However, they typically focus on manually crafted prompts and a restricted set of simple questions, limiting their accuracy and generalizability. Inspired by the human bias assessment process, we propose RoBGuard, a novel framework for enhancing LLMs to assess the risk of bias in RCTs. Specifically, RoBGuard integrates medical knowledge-enhanced question reformulation, multimodal document parsing, and multi-expert collaboration to ensure both completeness and accuracy. Additionally, to address the lack of suitable datasets, we introduce two new datasets: RoB-Item and RoB-Domain. Experimental results demonstrate RoBGuard’s effectiveness on the RoB-Item dataset, outperforming existing methods.

Retrieval-augmented generation (RAG) effectively addresses issues of static knowledge and hallucination in large language models. Existing studies mostly focus on question scenarios with clear user intents and concise answers. However, it is prevalent that users issue broad, open-ended queries with diverse sub-intents, for which they desire rich and long-form answers covering multiple relevant aspects. To tackle this important yet underexplored problem, we propose a novel RAG framework, namely RichRAG. It includes a sub-aspect explorer to identify potential sub-aspects of input questions, a multi-faceted retriever to build a candidate pool of diverse external documents related to these sub-aspects, and a generative list-wise ranker, which is a key module to provide the top-k most valuable documents for the final generator. These ranked documents sufficiently cover various query aspects and are aware of the generator’s preferences, hence incentivizing it to produce rich and comprehensive responses for users. The training of our ranker involves a supervised fine-tuning stage to ensure the basic coverage of documents, and a reinforcement learning stage to align downstream LLM’s preferences to the ranking of documents. Experimental results on two publicly available datasets prove that our framework effectively and efficiently provides comprehensive and satisfying responses to users.

The recent surge in jailbreaking attacks has revealed significant vulnerabilities in Large Language Models (LLMs) when exposed to malicious inputs. While various defense strategies have been proposed to mitigate these threats, there has been limited research into the underlying mechanisms that make LLMs vulnerable to such attacks. In this study, we suggest that the self-safeguarding capability of LLMs is linked to specific activity patterns within their representation space. Although these patterns have little impact on the semantic content of the generated text, they play a crucial role in shaping LLM behavior under jailbreaking attacks. Our findings demonstrate that these patterns can be detected with just a few pairs of contrastive queries. Extensive experimentation shows that the robustness of LLMs against jailbreaking can be manipulated by weakening or strengthening these patterns. Further visual analysis provides additional evidence for our conclusions, providing new insights into the jailbreaking phenomenon. These findings highlight the importance of addressing the potential misuse of open-source LLMs within the community.

Implicitly abusive language (IAL), unlike its explicit counterpart, lacks overt slurs or unambiguously offensive keywords, such as “bimbo” or “scum”, making it challenging to detect and mitigate. While current research predominantly focuses on explicitly abusive language, the subtler and more covert forms of IAL remain insufficiently studied. The rapid advancement and widespread adoption of large language models (LLMs) have opened new possibilities for various NLP tasks, but their application to IAL detection has been limited. We revisit three very recent challenging datasets of IAL and investigate the potential of LLMs to enhance the detection of IAL in English through zero-shot and few-shot prompting approaches. We evaluate the models’ capabilities in classifying sentences directly as either IAL or benign, and in extracting linguistic features associated with IAL. Our results indicate that classifiers trained on features extracted by advanced LLMs outperform the best previously reported results, achieving near-human performance.

Cosine similarity is widely used to measure the similarity between two embeddings, while interpretations based on angle and correlation coefficient are common. In this study, we focus on the interpretable axes of embeddings transformed by Independent Component Analysis (ICA), and propose a novel interpretation of cosine similarity as the sum of semantic similarities over axes. The normalized ICA-transformed embeddings exhibit sparsity, enhancing the interpretability of each axis, and the semantic similarity defined by the product of the components represents the shared meaning between the two embeddings along each axis. The effectiveness of this approach is demonstrated through intuitive numerical examples and thorough numerical experiments. By deriving the probability distributions that govern each component and the product of components, we propose a method for selecting statistically significant axes.

Recently, directly using large language models (LLMs) has been shown to be the most reliable method to evaluate QA models. However, it suffers from limited interpretability, high cost, and environmental harm. To address these, we propose to use soft exact match (EM) with entity-driven answer set expansion. Our approach expands the gold answer set to include diverse surface forms, based on the observation that the surface forms often follow particular patterns depending on the entity type. The experimental results show that our method outperforms traditional evaluation methods by a large margin. Moreover, the reliability of our evaluation method is comparable to that of LLM-based ones, while offering the benefits of high interpretability and reduced environmental harm.

Large language models (LLMs) are naturally suitable for Chinese spelling check (CSC) task in few-shot scenarios due to their powerful semantic understanding and few-shot learning capabilities. Recent CSC research has begun to use LLMs as foundational models. However, most current datasets are primarily focused on errors generated during the text generation process, with little attention given to errors occurring in the modal conversion process. Furthermore, existing LLM-based CSC methods often rely on fixed prompt samples, which limits the performance of LLMs. Therefore, we propose a framework named RagID (Retrieval-Augment Generation and Iterative Discriminator Strategy). By utilizing semantic-based similarity search and an iterative discriminator mechanism, RagID can provide well-chosen prompt samples and reduce over-correction issues in LLM-based CSC. RagID demonstrates excellent effectiveness in few-shot scenarios. We conducted comprehensive experiments, and the results show that RagID achieves the best performance on dataset that include data from multiple domains and dataset containing modal conversion spelling errors. The dataset and method are available online.