COLING 2025 Papers

Despite the impressive capabilities of large language models (LLMs) in aspect-based sentiment analysis (ABSA), the role of syntactic information remains underexplored in LLMs. Syntactic structures are known to be crucial for capturing aspect-opinion relationships. To explore whether LLMs can effectively leverage syntactic information to improve ABSA performance, we propose a novel multi-step reasoning framework, the Syntax-Opinion-Sentiment Reasoning Chain (Syn-Chain). Syn-Chain sequentially analyzes syntactic dependencies, extracts opinions, and classifies sentiment. We introduce Syn-Chain into LLMs via zero-shot prompting, and results show that Syn-Chain significantly enhances ABSA performance, though smaller LLM exhibit weaker performance. Furthermore, we enhance smaller LLMs via distillation using GPT-3.5-generated Syn-Chain responses, achieving state-of-the-art ABSA performance. Our findings highlight the importance of syntactic information for improving LLMs in ABSA and offer valuable insights for future research.

We present the first dataset, an annotation scheme, discourse analysis, and baseline experiments on argumentation and domain content types in scholarly articles on political science, specifically on the theory of International Relations (IR). The dataset comprises over 1 600 sentences stemming from three foundational articles on Neo-Realism, Liberalism, and Constructivism. We show that our annotation scheme enables educationally-relevant insight into the scholarly IR discourse and that state-of-the-art classifiers, while effective in distinguishing basic argumentative elements (Claims and Support/Attack relations) reaching up to 0.97 micro F1 , require domain-specific training and fine-tuning on the more fine-grained tasks of relation and content type prediction.

An end-to-end argument mining (AM) pipeline takes a text as input and provides its argumentative structure as output by identifying and classifying the argument units and argument relations in the text. In this work, we approach AM using fine-tuned large language models (LLMs). We model the three main sub-tasks of the AM pipeline, as well as their joint formulation, as text generation tasks. We fine-tune eight popular quantized and non-quantized LLMs – LLaMA-3, LLaMA-3.1, Gemma-2, Mistral, Phi-3, Qwen-2 – which are among the most capable open-weight models, on the benchmark PE, AbstRCT, and CDCP datasets that represent diverse data sources. Our approach achieves state-of-the-art results across all AM sub-tasks and datasets, showing significant improvements over previous benchmarks.

Keyword Extraction (KE) is essential in Natural Language Processing (NLP) for identifying key terms that represent the main themes of a text, and it is vital for applications such as information retrieval, text summarisation, and document classification. Despite the development of various KE methods — including statistical approaches and advanced deep learning models — evaluating their effectiveness remains challenging. Current evaluation metrics focus on keyword quality, balance, and overlap with annotations from authors and professional indexers, but neglect real-world information retrieval needs. This paper introduces a novel evaluation method designed to overcome this limitation by using real query data from Google Trends and can be used with both supervised and unsupervised KE approaches. We applied this method to three popular KE approaches (YAKE, RAKE and KeyBERT) and found that KeyBERT was the most effective in capturing users’ top queries, with RAKE also showing surprisingly good performance. The code is open-access and publicly available.

The swift progress and widespread acceptance of artificial intelligence (AI) systems highlight a pressing requirement to comprehend both the capabilities and potential risks associated with AI. Given the linguistic complexity, cultural richness, and underrepresented status of Arabic in AI research, there is a pressing need to focus on Large Language Models (LLMs) performance and safety for Arabic related tasks. Despite some progress in their development, there is a lack of comprehensive trustworthiness evaluation benchmarks which presents a major challenge in accurately assessing and improving the safety of LLMs when prompted in Arabic. In this paper, we introduce AraTrust, the first comprehensive trustworthiness benchmark for LLMs in Arabic. AraTrust comprises 522 human-written multiple-choice questions addressing diverse dimensions related to truthfulness, ethics, privacy, illegal activities, mental health, physical health, unfairness, and offensive language. We evaluated a set of LLMs against our benchmark to assess their trustworthiness. GPT-4 was the most trustworthy LLM, while open-source models, particularly AceGPT 7B and Jais 13B, struggled to achieve a score of 60% in our benchmark. The benchmark dataset is publicly available at https://huggingface.co/datasets/asas-ai/AraTrust

Arabic, with its rich diversity of dialects, remains significantly underrepresented in Large Language Models, particularly in dialectal variations. We address this gap by introducing seven synthetic datasets in dialects alongside Modern Standard Arabic (MSA), created using Machine Translation (MT) combined with human post-editing. We present AraDiCE, a benchmark for Arabic Dialect and Cultural Evaluation. We evaluate LLMs on dialect comprehension and generation, focusing specifically on low-resource Arabic dialects. Additionally, we introduce the first-ever fine-grained benchmark designed to evaluate cultural awareness across the Gulf, Egypt, and Levant regions, providing a novel dimension to LLM evaluation. Our findings demonstrate that while Arabic-specific models like Jais and AceGPT outperform multilingual models on dialectal tasks, significant challenges persist in dialect identification, generation, and translation. This work contributes ≈45K post-edited samples, a cultural benchmark, and highlights the importance of tailored training to improve LLM performance in capturing the nuances of diverse Arabic dialects and cultural contexts. We have released the dialectal translation models and benchmarks developed in this study (https://huggingface.co/datasets/QCRI/AraDiCE)

Many languages lack broad-coverage, semantically annotated lexical resources, which limits empirical research on lexical semantics for these languages. In this paper, we report on how we automatically enriched the French Wiktionnary with general semantic classes, known as supersenses, using a limited amount of manually annotated data. We trained a classifier combining sense definition classification and sense exemplars classification. The resulting resource, with an evaluated supersense accuracy of nearly 85% (92% for hypersenses), is used in a case study illustrating how such an semantically enriched resource can be leveraged to empirically test linguistic hypotheses about the lexicon, on a large scale.

In this paper, we investigate the role of attention heads in Context-aware Machine Translation models for pronoun disambiguation in the English-to-German and English-to-French language directions. We analyze their influence by both observing and modifying the attention scores corresponding to the plausible relations that could impact a pronoun prediction. Our findings reveal that while some heads do attend the relations of interest, not all of them influence the models’ ability to disambiguate pronouns. We show that certain heads are underutilized by the models, suggesting that model performance could be improved if only the heads would attend one of the relations more strongly. Furthermore, we fine-tune the most promising heads and observe the increase in pronoun disambiguation accuracy of up to 5 percentage points which demonstrates that the improvements in performance can be solidified into the models’ parameters.

Implicit hate speech detection is challenging due to its subjectivity and context dependence, with existing models often struggling in outof-domain scenarios. We propose CONELA, a novel data refinement strategy that enhances model performance and generalization by integrating human annotation agreement with model training dynamics. By removing both easy and hard instances from the model’s perspective, while also considering whether humans agree or disagree and retaining ambiguous cases crucial for out-of-distribution generalization, CONELA consistently improves performance across multiple datasets and models. We also observe significant improvements in F1 scores and cross-domain generalization with the use of our CONELA strategy. Addressing data scarcity in smaller datasets, we introduce a weighted loss function and an ensemble strategy incorporating disagreement maximization, effectively balancing learning from limited data. Our findings demonstrate that refining datasets by integrating both model and human perspectives significantly enhances the effectiveness and generalization of implicit hate speech detection models. This approach lays a strong foundation for future research on dataset refinement and model robustness.

The meanings and relationships of words shift over time. This phenomenon is referred to as semantic shift. Research focused on understanding how semantic shifts occur over multiple time periods is essential for gaining a detailed understanding of semantic shifts. However, detecting change points only between adjacent time periods is insufficient for analyzing detailed semantic shifts, and using BERT-based methods to examine word sense proportions incurs a high computational cost. To address those issues, we propose a simple yet intuitive framework for how semantic shifts occur over multiple time periods by utilizing similarity matrices based on word embeddings. We calculate diachronic word similarity matrices using fast and lightweight word embeddings across arbitrary time periods, making it deeper to analyze continuous semantic shifts. Additionally, by clustering the resulting similarity matrices, we can categorize words that exhibit similar behavior of semantic shift in an unsupervised manner.

Readability-controlled text simplification (RCTS) rewrites texts to lower readability levels while preserving their meaning. RCTS models often depend on parallel corpora with readability annotations on both source and target sides. Such datasets are scarce and difficult to curate, especially at the sentence level. To reduce reliance on parallel data, we explore using instruction-tuned large language models for zero-shot RCTS. Through automatic and manual evaluations, we examine: (1) how different types of contextual information affect a model’s ability to generate sentences with the desired readability, and (2) the trade-off between achieving target readability and preserving meaning. Results show that all tested models struggle to simplify sentences (especially to the lowest levels) due to models’ limitations and characteristics of the source sentences that impede adequate rewriting. Our experiments also highlight the need for better automatic evaluation metrics tailored to RCTS, as standard ones often misinterpret common simplification operations, and inaccurately assess readability and meaning preservation.

Biomedical Terminology Normalization aims to identify the standard term in a specified termbase for non-standardized mentions from social media or clinical texts, employing the mainstream “Recall and Re-rank” framework. Instead of the traditional pretraining-finetuning paradigm, we would like to explore the possibility of accomplishing this task through a tuning-free paradigm using powerful Large Language Models (LLMs), hoping to address the costs of re-training due to discrepancies of both standard termbases and annotation protocols. Another major obstacle in this task is that both mentions and terms are short texts. Short texts contain an insufficient amount of information that can introduce ambiguity, especially in a biomedical context. Therefore, besides using the advanced embedding model, we implement a Retrieval-Augmented Generation (RAG) based knowledge card generation module. This module introduces an LLM agent that expands the short texts into accurate, harmonized, and more informative descriptions using a search engine and a domain knowledge base. Furthermore, we present an innovative tuning-free agent collaboration framework for the biomedical terminology normalization task in social media. By leveraging the internal knowledge and the reasoning capabilities of LLM, our framework conducts more sophisticated recall, ranking and re-ranking processes with the collaboration of different LLM agents. Experimental results across multiple datasets indicate that our approach exhibits competitive performance. We release our code and data on the github repository JOHNNY-fans/RankNorm.

Tabular prediction, a well-established problem in machine learning, has consistently garnered significant research attention within academia and industry. Recently, with the rapid development of large language models (LLMs), there has been increasing exploration of how to apply LLMs to tabular prediction tasks. Many existing methods, however, typically rely on extensive pre-training or fine-tuning of LLMs, which demands considerable computational resources. To avoid this, we propose a retrieval-based approach that utilizes the powerful capabilities of LLMs in representation, comprehension, and inference. Our approach eliminates the need for training any modules or performing data augmentation, depending solely on information from target dataset. Experimental results reveal that, even without specialized training for tabular data, our method exhibits strong predictive performance on tabular prediction task, affirming its practicality and effectiveness.

Dialogue policy trains an agent to select dialogue actions frequently implemented via deep reinforcement learning (DRL). The model-based reinforcement methods built a world model to generate simulated data to alleviate the sample inefficiency. However, traditional world model methods merely consider one-step dialogues, leading to an inaccurate environmental simulation. Furthermore, different users may have different intention preferences, while most existing studies lack consideration of the intention-preferences causal relationship. This paper proposes a novel framework for dialogue policy learning named MCA, implemented through model-based reinforcement learning with automatically constructed causal chains. The MCA model utilizes an autoregressive Transformer to model dialogue trajectories, enabling a more accurate simulation of the environment. Additionally, it constructs a causal chains module that outputs latent preference distributions for intention-action pairs, thereby elucidating the relationship between user intentions and agent actions. The experimental results show that MCA can achieve state-of-the-art performances on three dialogue datasets over the compared dialogue agents, highlighting its effectiveness and robustness.

In sectors in where data quality is critical, like finance and healthcare, it is crucial to have confidence in not only the outputs generated by retrieval-augmented generation (RAG) models but also the process followed by the model while arriving at the output. Existing methods, such as hallucination detection and input-output entailment measurements, fail to capture the model’s internal state during answer generation. This paper introduces a novel approach to predict the correctness of the generated answer by modeling the model’s uncertainty on quantified perturbations of input. Extensive experiments across multiple large language models (LLMs) demonstrate that our approach quantifies RAG robustness by aligning predictions with ground truth with a Avg.Mean Square Error (MSE) 0.002 while offering flexibility for diverse qualitative metrics.

Ensuring that Large Language Models (LLMs) generate text representative of diverse sub-populations is essential, particularly when key concepts related to under-represented groups are scarce in the training data. We address this challenge with a novel clustering-based active learning framework, enhanced with knowledge distillation. The proposed framework transforms the intermediate outputs of the learner model, enabling effective active learning for generative tasks for the first time. Integration of clustering and knowledge distillation yields more representative models without prior knowledge of underlying data distribution and overbearing human efforts. We validate our approach in practice through case studies in counter-narration and style transfer. We construct two new datasets in tandem with model training, showing a performance improvement of 2%–10% over baseline models. Our results also show more consistent performance across various data subgroups and increased lexical diversity, underscoring our model’s resilience to skewness in available data. Further, our results show that the data acquired via our approach improves the performance of secondary models not involved in the learning loop, showcasing practical utility of the framework.

Multi-modal sarcasm detection aims to identify whether a given image-text pair is sarcastic. The pivotal factor of the task lies in accurately capturing incongruities from different modalities. Although existing studies have achieved impressive success, they primarily committed to fusing the textual and visual information to establish cross-modal correlations, overlooking the significance of original unimodal incongruity information at the text-level and image-level. Furthermore, the utilized fusion strategies of cross-modal information neglected the effect of inherent ambiguity within text and image modalities on multimodal fusion. To overcome these limitations, we propose a novel Ambiguity-aware Multi-level Incongruity Fusion Network (AMIF) for multi-modal sarcasm detection. Our method involves a multi-level incongruity learning module to capture the incongruity information simultaneously at the text-level, image-level and cross-modal-level. Additionally, an ambiguity-based fusion module is developed to dynamically learn reasonable weights and interpretably aggregate incongruity features from different levels. Comprehensive experiments conducted on a publicly available dataset demonstrate the superiority of our proposed model over state-of-the-art methods.

Machine unlearning aims to efficiently eliminate the influence of specific training data, known as the forget set, from the model. However, existing unlearning methods for Large Language Models (LLMs) face a critical challenge: they rely solely on negative feedback to suppress responses related to the forget set, which often results in nonsensical or inconsistent outputs, diminishing model utility and posing potential privacy risks. To address this limitation, we propose a novel approach called Alternate Preference Optimization (AltPO), which combines negative feedback with in-domain positive feedback on the forget set. Additionally, we introduce new evaluation metrics to assess the quality of responses related to the forget set. Extensive experiments show that our approach not only enables effective unlearning but also avoids undesirable model behaviors while maintaining overall model performance.

Open-Domain Question Answering (ODQA) systems often struggle with the quality of retrieved passages, which may contain conflicting information and be misaligned with the reader’s needs. Existing retrieval methods aim to gather relevant passages but often fail to prioritize consistent and useful information for the reader. In this paper, we introduce a novel Reader-Centered Passage Selection (R-CPS) method, which enhances the performance of the retrieve-then-read pipeline by re-ranking and clustering passages from the reader’s perspective. Our method re-ranks passages based on the reader’s prediction probability distribution and clusters passages according to the predicted answers, prioritizing more useful and relevant passages to the top and reducing inconsistent information. Experiments on ODQA datasets demonstrate the effectiveness of our approach in improving the quality of evidence passages under zero-shot settings.