COLING 2025 Papers

Large language models (LLMs) exhibit remarkable capabilities in understanding and generating natural language. However, these models can inadvertently memorize private information, posing significant privacy risks. This study addresses the challenge of enabling LLMs to protect specific individuals’ private data without the need for complete retraining. We propose RETURN, a Real-world pErsonal daTa UnleaRNing dataset, comprising 2,492 individuals from Wikipedia with associated QA pairs, to evaluate machine unlearning (MU) methods for protecting personal data in a realistic scenario. Additionally, we introduce the Name-Aware Unlearning Framework (NAUF) for Privacy Protection, which enables the model to learn which individuals’ information should be protected without affecting its ability to answer questions related to other unrelated individuals. Our extensive experiments demonstrate that NAUF achieves a state-of-the-art average unlearning score, surpassing the best baseline method by 5.65 points, effectively protecting target individuals’ personal data while maintaining the model’s general capabilities.

Small language models (SLMs) are more efficient, cost-effective, and customizable than large language models (LLMs), though they often underperform in specific areas like reasoning. Past methods for enhancing SLMs’ reasoning, such as supervised fine-tuning and distillation, often depend on costly external signals, resulting in SLMs being overly confident with limited supervision signals, thus limiting their abilities. Therefore, this study enables SLMs to learn to reason from self-iterative feedback. By combining odds ratio preference optimization (ORPO), we fine-tune and align SLMs using positive and negative signals generated by themselves. Additionally, we introduce process supervision for rewards in preference alignment by sampling-based inference simulation and process reward models. Compared to Supervised Fine-Tuning (SFT), our method improves the performance of Gemma-2B by 12.43 (Acc) on GSM8K and 3.95 (Pass@1) on MBPP. Furthermore, the proposed method also demonstrated superior out-of-domain generalization capabilities on MMLU_Math and HumanEval.

This position paper investigates the potential of integrating insights from language impairment research and its clinical treatment to develop human-inspired learning strategies and evaluation frameworks for language models (LMs). We inspect the theoretical underpinnings underlying some influential linguistically motivated training approaches derived from neurolinguistics and, particularly, aphasiology, aimed at enhancing the recovery and generalization of linguistic skills in aphasia treatment, with a primary focus on those targeting the syntactic domain. We highlight how these insights can inform the design of rigorous assessments for LMs, specifically in their handling of complex syntactic phenomena, as well as their implications for developing human-like learning strategies, aligning with efforts to create more sustainable and cognitively plausible natural language processing (NLP) models.

Large Language Models (LLMs) have demonstrated powerful performance in sequential recommendation due to their robust language modeling and comprehension capabilities. In such paradigms, the item texts of interaction sequences are formulated as sentences and LLMs are utilized to learn language representations or directly generate target item texts by incorporating instructions. Despite their promise, these methods solely focus on modeling the mapping from sequential texts to target items, neglecting the relationship between the items in an interaction sequence. This results in a failure to learn the transition patterns between items, which reflect the dynamic change in user preferences and are crucial for predicting the next item. To tackle this issue, we propose a novel framework for mapping the sequential item texts to the sequential item IDs, named ST2SI. Specifically, we first introduce multi-query input and item linear projection (ILP) to model the conditional probability distribution of items. Then, we further propose ID alignment to address misalignment between item texts and item IDs by instruction tuning. Finally, we propose efficient ILP tuning to adapt flexibly to different scenarios, requiring only training a linear layer to achieve competitive performance. Extensive experiments on six real-world datasets show our approach outperforms the best baselines by 7.33% in NDCG@10, 4.65% in Recall@10, and 8.42% in MRR.

Implicit hate speech presents a significant challenge for automatic detection systems due to its subtlety and ambiguity. Traditional models trained using empirical risk minimization (ERM) often rely on correlations between class labels and spurious attributes, which leads to poor performance on data lacking these correlations. In this paper, we propose a novel approach using causality-guided contrastive learning (CCL) to enhance the generalizability of implicit hate speech detection. Since ERM tends to identify spurious attributes, CCL works by aligning the representations of samples with the same class but opposite spurious attributes, identified through ERM’s inference failure. This method reduces the model’s reliance on spurious correlations, allowing it to learn more robust features and handle diverse, nuanced contexts better. Our extensive experiments on multiple implicit hate speech datasets show that our approach outperforms current state-of-the-art methods in cross-domain generalization.

Recent state-of-the-art authorship attribution methods learn authorship representations of text in a latent, uninterpretable space, which hinders their usability in real-world applications. We propose a novel approach for interpreting learned embeddings by identifying representative points in the latent space and leveraging large language models to generate informative natural language descriptions of the writing style associated with each point. We evaluate the alignment between our interpretable and latent spaces and demonstrate superior prediction agreement over baseline methods. Additionally, we conduct a human evaluation to assess the quality of these style descriptions and validate their utility in explaining the latent space. Finally, we show that human performance on the challenging authorship attribution task improves by +20% on average when aided with explanations from our method.

Recent advancements in natural language processing, driven by Large Language Models (LLMs), have significantly improved text comprehension, enabling these models to handle complex tasks with greater efficiency. A key feature of LLMs is their ability to engage in contextual learning, which allows them to understand and apply instructions given in natural language to new scenarios without requiring additional training. This capability is particularly valuable in social media, where LLMs can be crucial in addressing challenges in explainable sexism detection. We hypothesize that by leveraging contextual learning capabilities, LLMs can provide clear, explainable insights into why certain content is flagged as problematic, thus enhancing transparency in the sexism detection process. To this end, we propose a Reinforcement Learning from Human Feedback (RLHF) based fine-tuning framework for sexism detection. We studied two well-known LLMs, Mistral-7B and LLaMA-3-8B, in zero-shot, supervised fine-tuning, and RLHF scenarios to conclude the superior ability of LLMs in sexism detection. The experimental results reported in this work, based on three tasks of Explainable Detection of Online Sexism (EDOS), highlight the importance of RLHF for building explainable systems in online discourse. Furthermore, we found that the LLaMA-3-8B model achieves the best results using the RLHF approach, scoring 0.8681 on Task A (binary sexism detection), 0.6829 on Task B (category classification of sexism), and 0.4722 on Task C (fine-grained sexism vectors) test sets.

Recently, increasing attention has been focused on improving the ability of Large Language Models (LLMs) to perform complex reasoning. Advanced methods, such as Chain-of-Thought (CoT) and its variants, are found to enhance their reasoning skills by designing suitable prompts or breaking down complex problems into more manageable sub-problems. However, little concentration has been put on exploring the reasoning process, i.e., we discovered that most methods resort to Direct Reasoning (DR) and disregard Indirect Reasoning (IR). This can make LLMs difficult to solve IR tasks, which are often encountered in the real world. To address this issue, we propose a Direct-Indirect Reasoning (DIR) method, which considers DR and IR as multiple parallel reasoning paths that are merged to derive the final answer. We stimulate LLMs to implement IR by crafting prompt templates incorporating the principles of contrapositive and contradiction. These templates trigger LLMs to assume the negation of the conclusion as true, combine it with the premises to deduce a conclusion, and utilize the logical equivalence of the contrapositive to enhance their comprehension of the rules used in the reasoning process. Our DIR method is simple yet effective and can be straightforwardly integrated with existing variants of CoT methods. Experimental results on four datasets related to logical reasoning and mathematic proof demonstrate that our DIR method, when combined with various baseline methods, significantly outperforms all the original methods.

With the advent of Large Language Models (LLMs), generating rule-based data for real-world applications has become more accessible. Due to the inherent ambiguity of natural language and the complexity of rule sets, especially in long contexts, LLMs often struggle to follow all specified rules, frequently omitting at least one. To enhance the reasoning and understanding of LLMs on long and complex contexts, we propose a novel prompting strategy Multi-Lingual Prompt, namely MLPrompt, which automatically translates the error-prone rule that an LLM struggles to follow into another language, thus drawing greater attention to it. Experimental results on public datasets across various tasks have shown MLPrompt can outperform state-of-the-art prompting methods such as Chain of Thought, Tree of Thought, and Self-Consistency. Additionally, we introduce a framework integrating MLPrompt with an auto-checking mechanism for structured data generation, with a specific case study in text-to-MIP instances. Further, we extend the proposed framework for text-to-SQL to demonstrate its generation ability towards structured data synthesis.

Large Language Models (LLMs) have achieved outstanding performance across various NLP tasks. Grammatical Error Correction (GEC) is a task aiming at automatically correcting grammatical errors in text, but it encounters a severe shortage of annotated data. Researchers have tried to make full use of the generalization capabilities of LLMs and prompt them to correct erroneous sentences, which however results in unexpected over-correction issues. In this paper, we rethink the role of LLMs in GEC tasks and propose a method, namely TypeDA, considering LLMs as the annotators for type-aware data augmentation in GEC tasks. Different from the existing data augmentation methods, our method prevents in-distribution corruption and is able to generate sentences with multi-granularity error types. Our experiments verify that our method can generally improve the GEC performance of different backbone models with only a small amount of augmented data. Further analyses verify the high consistency and diversity of the pseudo data generated via our method.

Event Relation Extraction (ERE) aims to extract various types of relations between different events within texts. Although Large Language Models (LLMs) have demonstrated impressive capabilities in many natural language processing tasks, existing ERE methods based on LLMs still face three key challenges: (1) Time Inefficiency: The existing pairwise method of combining events and determining their relations is time-consuming for LLMs. (2) Low Coverage: When dealing with numerous events in a document, the limited generation length of fine-tuned LLMs restricts the coverage of their extraction results. (3) Lack of Rationale: Essential rationales concerning the results that could enhance the reasoning ability of the model are overlooked. To address these challenges, we propose LLMERE, an LLM-based approach with rationales for the ERE task. LLMERE transforms ERE into a question-and-answer task that may have multiple answers. By extracting all events related to a specified event at once, LLMERE reduces time complexity from O(n2) to O(n), compared to the pairwise method. Subsequently, LLMERE enhances the coverage of extraction results by employing a partitioning strategy that highlights only a portion of the events in the document at a time. In addition to the extracted results, LLMERE is also required to generate corresponding rationales/reasons behind them, in terms of event coreference information or transitive chains of event relations. Experimental results on three widely used datasets show that LLMERE achieves significant improvements over baseline methods.

Extreme Multi-label Text Classification (XMC) entails selecting the most relevant labels for an instance from a vast label set. Extreme Zero-shot XMC (EZ-XMC) extends this challenge by operating without annotated data, relying only on raw text instances and a predefined label set, making it particularly critical for addressing cold-start problems in large-scale recommendation and categorization systems. State-of-the-art methods, such as MACLR and RTS, leverage lightweight bi-encoders but rely on suboptimal pseudo labels for training, such as document titles (MACLR) or document segments (RTS), which may not align well with the intended tagging or categorization tasks. On the other hand, LLM-based approaches, like ICXML, achieve better label-instance alignment but are computationally expensive and impractical for real-world EZ-XMC applications due to their heavy inference costs. In this paper, we introduce LMTX (Large language Model as Teacher for eXtreme classification), a novel framework that bridges the gap between these two approaches. LMTX utilizes an LLM to identify high-quality pseudo labels during training, while employing a lightweight bi-encoder for efficient inference. This design eliminates the need for LLMs at inference time, offering the benefits of improved label alignment without sacrificing computational efficiency. Our approach achieves superior performance and efficiency over both LLM and non-LLM based approaches, establishing a new state-of-the-art in EZ-XMC.

Advancements in language models (LMs) have sparked interest in exploring their potential as knowledge bases (KBs) due to their high capability for storing huge amounts of factual knowledge and semantic understanding. However, existing studies face challenges in quantifying the extent of large-scale knowledge packed into LMs and lack systematic studies on LMs’ structured reasoning capabilities over the infused knowledge. Addressing these gaps, our research investigates whether LMs can effectively act as large-scale KBs after training over an expansive set of world knowledge triplets via addressing the following three crucial questions: (1) How do LMs of different sizes perform at storing world knowledge of different frequencies in a large-scale KB? (2) How flexible are these LMs in recalling the stored knowledge when prompted with natural language queries? (3) After training on the abundant world knowledge, can LMs additionally gain the ability to reason over such information to infer new facts? Our findings indicate that while medium-scaled LMs hold promise as world knowledge bases capable of storing and responding with flexibility, enhancements in their reasoning capabilities are necessary to fully realize their potential.

This paper explores whether language models can effectively resolve the complex binding patterns of the Mandarin Chinese reflexive ziji, which are constrained by both syntactic and semantic factors. We construct a dataset of 320 synthetic sentences using templates and examples from syntactic literature, along with 360 natural sentences from the BCC corpus. Evaluating 21 language models against this dataset and comparing their performance to judgments from native Mandarin speakers, we find that none of the models consistently replicates human-like judgments. The results indicate that existing language models tend to rely heavily on sequential cues, though not always favoring the closest strings, and often overlooking subtle semantic and syntactic constraints. They tend to be more sensitive to noun-related than verb-related semantics.

Large language models (LLMs) have exhibited impressive competence in various tasks, but their internal mechanisms on mathematical problems are still under-explored. In this paper, we study a fundamental question: how language models encode the value of numbers, a basic element in math. To study the question, we construct a synthetic dataset comprising addition problems and utilize linear probes to read out input numbers from the hidden states. Experimental results support the existence of encoded number values in LLMs on different layers, and these values can be extracted via linear probes. Further experiments show that LLMs store their calculation results in a similar manner, and we can intervene the output via simple vector additions, proving the causal connection between encoded numbers and language model outputs. Our research provides evidence that LLMs encode the value of numbers linearly, offering insights for better exploring, designing, and utilizing numeric information in LLMs.

There has been a surge of interest regarding language adaptation of Large Language Models (LLMs) to enhance the processing of texts in low-resource languages. While traditional language models have seen extensive research on language transfer, modern LLMs still necessitate further explorations in language adaptation. In this paper, we present a systematic review of the language adaptation process for LLMs, including vocabulary expansion, continued pre-training, and instruction fine-tuning, which focuses on empirical studies conducted on LLaMA2 and discussions on various settings affecting the model’s capabilities. This study provides helpful insights covering the entire language adaptation process, and highlights the compatibility and interactions between different steps, offering researchers a practical guidebook to facilitate the effective adaptation of LLMs across different languages.

Emotion recognition in conversation (ERC), the task of discerning human emotions for each utterance within a conversation, has garnered significant attention in human-computer interaction systems. Previous ERC studies focus on speaker-specific information that predominantly stems from relationships among utterances, which lacks sufficient information around conversations. Recent research in ERC has sought to exploit pre-trained large language models (LLMs) with speaker modelling to comprehend emotional states. Although these methods have achieved the encouraging results, the extracted speaker-specific information struggles to indicate emotional dynamics. In this paper, motivated by the fact that speaker characteristics play a crucial role and LLMs have rich world knowledge, we present LaERC-S, a novel framework that stimulates LLMs to explore speaker characteristics involving the mental state and behavior of interlocutors, for accurate emotion predictions. To endow LLMs with these knowledge information, we adopt the two-stage learning to make the models reason speaker characteristics and track the emotion of the speaker in complex conversation scenarios. Extensive experiments on three benchmark datasets demonstrate the superiority of LaERC-S, reaching the new state-of-the-art.

We introduce LUCE, an advanced dynamic framework with an interactive dashboard for analysing opinionated text aiming to understand people-centred communication. The framework features computational modules of text classification and extraction explicitly designed for analysing different elements of opinions, e.g., sentiment/emotion, suggestion, figurative language, hate/toxic speech, and topics. We designed the framework using a modular architecture, allowing scalability and extensibility with the aim of supporting other NLP tasks in subsequent versions. LUCE comprises trained models, python-based APIs, and a user-friendly dashboard, ensuring an intuitive user experience. LUCE has been validated in a relevant environment, and its capabilities and performance have been demonstrated through initial prototypes and pilot studies.

Word Sense Disambiguation (WSD) is a fundamental task critical for accurate semantic understanding. Conventional training strategies usually only consider predefined senses for target words and learn each of them from relatively limited instances, neglecting the influence of similar ones. To address these problems, we propose the method of Learning to Rank Senses (LTRS) to enhance the task. This method helps a model learn to represent and disambiguate senses from a broadened range of instances via ranking an expanded list of sense definitions. By employing LTRS, our model achieves a SOTA F1 score of 79.6% in Chinese WSD and exhibits robustness in low-resource settings. Moreover, it shows excellent training efficiency, achieving faster convergence than previous methods. This provides a new technical approach to WSD and may also apply to the task for other languages.