COLING 2025 论文列表

The task of determining whether two texts are paraphrases has long been a challenge in NLP. However, the prevailing notion of paraphrase is often quite simplistic, offering only a limited view of the vast spectrum of paraphrase phenomena. Indeed, we find that evaluating models in a paraphrase dataset can leave uncertainty about their true semantic understanding. To alleviate this, we create PARAPHRASUS, a benchmark designed for multi-dimensional assessment, benchmarking and selection of paraphrase detection models. We find that paraphrase detection models under our fine-grained evaluation lens exhibit trade-offs that cannot be captured through a single classification dataset. Furthermore, PARAPHRASUS allows prompt calibration for different use cases, tailoring LLM models to specific strictness levels. PARAPHRASUS includes 3 challenges spanning over 10 datasets, including 8 repurposed and 2 newly annotated; we release it along with a benchmarking library at https://github.com/impresso/paraphrasus

As personality can be useful in many cases, such as better understanding people’s underlying contexts or providing personalized services, research has long focused on modeling personality from data. However, the development of personality detection models faces challenges due to the inherent latent and relative characteristics of personality, as well as the lack of annotated datasets. To address these challenges, our research focuses on methods that effectively exploit the inherent knowledge of Large Language Models (LLMs). We propose a novel approach that compares contrasting perspectives to better capture the relative nature of personality traits. In this paper, we introduce PADO (Personality-induced multi-Agent framework for Detecting OCEAN of the Big Five personality traits), the first LLM-based multi-agent personality detection framework. PADO employs personality-induced agents to analyze text from multiple perspectives, followed by a comparative judgment process to determine personality trait levels. Our experiments with various LLM models, from GPT-4o to LLaMA3-8B, demonstrate PADO’s effectiveness and generalizability, especially with smaller parameter models. This approach offers a more nuanced, context-aware method for personality detection, potentially improving personalized services and insights into digital behavior. We will release our codes.

The recent emergence of various large language models, which can be fine-tuned with minimal instruction data, has demonstrated impressive performance across various tasks. However, a phenomenon of forgetting occurs during life- long fine-tuning because training on new tasks interferes with the previously acquired knowl- edge. To mitigate catastrophic forgetting, con- ventional data replay methods achieve high per- formance, but at the cost of compromising data privacy and security. This paper introduces a dataless distribution replay approach for life- long fine-tuning. Concretely, the distribution distillation is applied to replay the output dis- tribution of the linear layers at previous task stages. The optimal solution for this distri- bution replay can be directly computed using the retained inner product matrix of the input data, thereby eliminating the need for previ- ous data. Additionally, Singular Value Decom- position (SVD) and module accumulation are employed to further enhance the performance of dataless distribution replay method. Finally, the evaluation is conducted in a lifelong fine- tuning scenario involving multiple tasks. The experimental results and analysis show that the proposed method achieves significant improve- ments compared to several strong lifelong fine- tuning methods.

Large language models (LLMs) have achieved notable success across various tasks but are hindered by their large size and high computational demands. Post-training pruning (PTP) offers a promising solution by reducing model size through parameter removal while preserving performance. However, current PTP methods perform optimally only within specific sparsity ranges. This paper presents two key findings: (1) Layerwise uniform sparsity is effective at low sparsity, while non-uniform sparsity excels at high levels; (2) Relative importance-based pruning works best at low sparsity, whereas Hessian-based weight reconstruction is superior at high sparsity. We design and conduct experiments to validate these findings. Based on these insights, we introduce OptiPrune, a robust pruning method effective across all sparsity levels. OptiPrune adapts non-uniform sparsity with adaptive deviation and employs a threshold to select the optimal pruning strategy. Empirical results across diverse datasets, architectures, and languages validate its performance and robustness. These findings provide valuable directions for future LLM pruning research. Our code and data are publicly available.

Complex events generally exhibit unforeseen, multifaceted, and multi-step developments, and cannot be well handled by existing closed-ended event forecasting methods, which are constrained by a limited answer space. In order to accelerate the research on complex event forecasting, we introduce OpenForecast, a large-scale open-ended dataset with two features: (1) OpenForecast defines three open-ended event forecasting tasks, enabling unforeseen, multifaceted, and multi-step forecasting. (2) OpenForecast collects and annotates a large-scale dataset from Wikipedia and news, including 43,419 complex events spanning from 1950 to 2024. Particularly, this annotation can be completed automatically without any manual annotation cost. Meanwhile, we introduce an automatic LLM-based Retrieval-Augmented Evaluation method (LRAE) for complex events, enabling OpenForecast to evaluate the ability of complex event forecasting of large language models. Finally, we conduct comprehensive human evaluations to verify the quality and challenges of OpenForecast, and the consistency between LEAE metric and human evaluation. OpenForecast and related codes will be publicly released.

The increased use of large language models (LLMs) across a variety of real-world applications calls for mechanisms to verify the fac- tual accuracy of their outputs. Difficulties lie in assessing the factuality of free-form responses in open domains. Also, different pa- pers use disparate evaluation benchmarks and measurements, which renders them hard to compare and hampers future progress. To mitigate these issues, we propose OpenFactCheck, a unified framework for building customized automatic fact-checking systems, benchmarking their accuracy, evaluating factuality of LLMs, and verifying claims in a document. OpenFactCheck consists of three modules: (i) CUSTCHECKER allows users to easily customize an automatic fact-checker and verify the factual correctness of documents and claims, (ii) LLMEVAL, a unified evaluation framework assesses LLM’s factuality ability from various perspectives fairly, and (iii) CHECKEREVAL is an extensible solution for gauging the reliability of automatic fact-checkers’ verification results using human-annotated datasets. Data and code are publicly available at https: //github.com/yuxiaw/openfactcheck.

Although Wikipedia is the largest multilingual encyclopedia, it remains inherently incomplete. There is a significant disparity in the quality of content between high-resource languages (HRLs, e.g., English) and low-resource languages (LRLs, e.g., Hindi), with many LRL articles lacking adequate information. To bridge these content gaps, we propose a lightweight framework to enhance knowledge equity between English and Hindi. In case the English Wikipedia page is not up-to-date, our framework extracts relevant information from external resources readily available (such as English books), and adapts it to align with Wikipedia’s distinctive style, including its neutral point of view (NPOV) policy, using in-context learning capabilities of large language models. The adapted content is then machine-translated into Hindi for integration into the corresponding Wikipedia articles. On the other hand, if the English version is comprehensive and up-to-date, the framework directly transfers knowledge from English to Hindi. Our framework effectively generates new content for Hindi Wikipedia sections, enhancing Hindi Wikipedia articles respectively by 65% and 62% according to automatic and human judgment-based evaluations.

Visual7W has been widely used in assessing multiple-choice visual question-answering (VQA) systems. This paper reports on a replicated human experiment on Visual7W with the aim of understanding the human-level performance of VQA. The replication was not entirely successful because human participants performed significantly worse when answering “where”, “when”, and “how” questions in compared to other question types. An error analysis discovered that the failure was a consequence of the non-deterministic distractors in Visual7W. GPT-4V was then evaluated using and was compared to the human-level performance. The results embody that, when evaluating models’ capacity on Visual7W, the performance is not necessarily the higher, the better.

Text-based reinforcement learning involves an agent interacting with a fictional environment using observed text and admissible actions in natural language to complete a task. Previous works have shown that agents can succeed in text-based interactive environments even in the complete absence of semantic understanding or other linguistic capabilities. The success of these agents in playing such games suggests that semantic understanding may not be important for the task. This raises an important question about the benefits of LMs in guiding the agents through the game states. In this work, we show that rich semantic understanding leads to efficient training of text-based RL agents. Moreover, we describe the occurrence of semantic degeneration as a consequence of inappropriate fine-tuning of language models in text-based reinforcement learning (TBRL). Specifically, we describe the shift in the semantic representation of words in the LM, as well as how it affects the performance of the agent in tasks that are semantically similar to the training games. These results may help develop better strategies to fine-tune agents in text-based RL scenarios.

Large language models (LLMs) have made significant advancements, but their increasing capabilities present serious risks of misuse, particularly in open-weight models where direct access to the model’s parameters is possible. Current safeguards, designed for closed-weight API models, are inadequate for open-weight models, as minimal fine-tuning can bypass these protections. Preserving the integrity of open-weight LLMs before deployment has thus become a critical challenge. We argue that these vulnerabilities stem from the overemphasis on maximizing the LLM’s log-likelihood during training, which amplifies data biases, especially with large datasets. To address these issues, we introduce Kahneman and Tversky’s Prospect Theoretic Integrity Preserving Alignment (KT-IPA), a framework that prioritizes maximizing generative utility rather than a singular optimization metric. This approach strengthens LLMs against misuse and weaponization while maintaining high performance, even after extensive fine-tuning. Our results demonstrate that integrating prospect theory into LLM training enhances robustness, security, and responsible innovation in this rapidly evolving field. Our codes are available on https://anonymous.4open.science/r/KT-IPA-40B7

Textual data augmentation (DA) is a prolific field of study where novel techniques to create artificial data are regularly proposed, and that has demonstrated great efficiency on small data settings, at least for text classification tasks. In this paper, we challenge those results, showing that classical data augmentation (which modify sentences) is simply a way of performing better fine-tuning, and that spending more time doing so before applying data augmentation negates its effect. This is a significant contribution as it answers several questions that were left open in recent years, namely : which DA technique performs best (all of them as long as they generate data close enough to the training set, as to not impair training) and why did DA show positive results (facilitates training of network). We further show that zero- and few-shot DA via conversational agents such as ChatGPT or LLama2 can increase performances, confirming that this form of data augmentation is preferable to classical methods.

Recent works have successfully applied Large Language Models (LLMs) to function modeling tasks. However, the reasons behind this success remain unclear. In this work, we propose a new evaluation framework to comprehensively assess LLMs’ function modeling abilities. By adopting a Bayesian perspective of function modeling, we discover that LLMs are relatively weak in understanding patterns in raw data, but excel at utilizing prior knowledge about the domain to develop a strong understanding of the underlying function. Our findings offer new insights about the strengths and limitations of LLMs in the context of function modeling.

We present a new method to detect anomalies in texts (in general: in sequences of any data), using language models, in a totally unsupervised manner. The method considers probabilities (likelihoods) generated by a language model, but instead of focusing on low-likelihood tokens, it considers a new metric defined in this paper: oddballness. Oddballness measures how “strange” a given token is according to the language model. We demonstrate in grammatical error detection tasks (a specific case of text anomaly detection) that oddballness is better than just considering low-likelihood events, if a totally unsupervised setup is assumed.

Recently, Multi-modal Entity Linking (MEL) has attracted increasing attention in the research community due to its significance in numerous multi-modal applications. Video, as a popular means of information transmission, has become prevalent in people’s daily lives. However, most existing MEL methods primarily focus on linking textual and visual mentions or offline videos’ mentions to entities in multi-modal knowledge bases, with limited efforts devoted to linking mentions within online video content. In this paper, we propose a task called Online Video Entity Linking (OVEL), aiming to establish connections between mentions in online videos and a knowledge base with high accuracy and timeliness. To facilitate the research works of (OVEL), we specifically concentrate on live delivery scenarios and construct a live delivery entity linking dataset called (LIVE). Besides, we propose an evaluation metric that considers robustness, timelessness, and accuracy. Furthermore, to effectively handle (OVEL) task, we leverage a memory block managed by a Large Language Model and retrieve entity candidates from the knowledge base to augment LLM performance on memory management. The experimental results prove the effectiveness and efficiency of our method.

Current keyword decision-making in sponsored search advertising relies on large static datasets, limiting automatic keyword setup and failing to adapt to real-time KPI metrics and product updates essential for effective advertising. In this paper, we propose On-the-fly Keyword Generation (OKG), an LLM agent-based method that dynamically monitors KPI changes and adapts keyword generation in real-time, realizing the strategy recommended by advertising platforms. Additionally, we introduce the first publicly accessible dataset containing real keyword data with its KPIs across diverse domains, providing a valuable resource for future research. Experimental results and ablation studies demonstrate the effectiveness of OKG, showing significant improvements across various metrics and emphasizing the importance of each component. We believe OKG not only pioneers the use of LLM agents in this research field but also offers practical value for thousands of advertisers to automate keyword generation in real-world applications.

The preservation and revitalization of endangered and extinct languages is a meaningful endeavor, conserving cultural heritage while enriching fields like linguistics and anthropology. However, these languages are typically low-resource, making their reconstruction labor-intensive and costly. This challenge is exemplified by Nüshu, a rare script historically used by Yao women in China for self-expression within a patriarchal society. To address this challenge, we introduce NüshuRescue, an AI-driven framework designed to train large language models (LLMs) on endangered languages with minimal data. NüshuRescue automates evaluation and expands target corpora to accelerate linguistic revitalization. As a foundational component, we developed NCGold, a 500-sentence Nüshu-Chinese parallel corpus, the first publicly available dataset of its kind. Leveraging GPT-4-Turbo, with no prior exposure to Nüshu and only 35 short examples from NCGold, NüshuRescue achieved 48.69% translation accuracy on 50 withheld sentences and generated NCSilver, a set of 98 newly translated modern Chinese sentences of varying lengths. In addition, we developed FastText-based and Seq2Seq models to further support research on Nüshu. NüshuRescue provides a versatile and scalable tool for the revitalization of endangered languages, minimizing the need for extensive human input. All datasets and code have been made publicly available at https://github.com/ivoryayang/NushuRescue.

The rapid expansion of online content has intensified the issue of information redundancy, underscoring the need for solutions that can identify genuinely new information. Despite this challenge, the research community has seen a decline in focus on novelty detection, particularly with the rise of large language models (LLMs). Additionally, previous approaches have relied heavily on human annotation, which is time-consuming, costly, and particularly challenging when annotators must compare a target document against a vast number of historical documents. In this work, we introduce NovAScore (Novelty Evaluation in Atomicity Score), an automated metric for evaluating document-level novelty. NovAScore aggregates the novelty and salience scores of atomic information, providing high interpretability and a detailed analysis of a document’s novelty. With its dynamic weight adjustment scheme, NovAScore offers enhanced flexibility and an additional dimension to assess both the novelty level and the importance of information within a document. Our experiments show that NovAScore strongly correlates with human judgments of novelty, achieving a 0.626 Point-Biserial correlation on the TAP-DLND 1.0 dataset and a 0.920 Pearson correlation on an internal human-annotated dataset.

In recent years, numerous studies have sought to understand the cognitive dynamics underlying language processing by modeling reading times and ERP amplitudes using computational metrics like surprisal. In the present paper, we examine the predictive power of surprisal, entropy, and a novel metric based on semantic similarity for N400 and P600. Our experiments, conducted with Mandarin Chinese materials, revealed three key findings: 1) expectancy plays a primary role for N400; 2) P600 also reflects the cognitive effort required to evaluate linguistic input semantically; and 3) during the time window of interest, information uncertainty influences the language processing the most. Our findings show how computational metrics that capture distinct cognitive dimensions can effectively address psycholinguistic questions.

Contextualized embeddings vary by context, even for the same token, and form a distribution in the embedding space. To analyze this distribution, we focus on the norm of the mean embedding and the variance of the embeddings. In this study, we first demonstrate that these values follow the well-known formula for variance in statistics and provide an efficient sequential computation method. Then, by observing embeddings from intermediate layers of several Transformer models, we found a strong trade-off relationship between the norm and the variance: as the mean embedding becomes closer to the origin, the variance increases. Furthermore, when the sets of token embeddings are treated as clusters, we show that the variance of the entire embedding set can theoretically be decomposed into the within-cluster variance and the between-cluster variance. We found experimentally that as the layers of Transformer models deepen, the embeddings move farther from the origin, the between-cluster variance relatively decreases, and the within-cluster variance relatively increases. These results are consistent with existing studies on the anisotropy of the embedding spaces across layers.