COLING 2025 Papers

The rise of Multi-modal Pre-training highlights the necessity for a unified Multi-Modal Knowledge Graph (MMKG) representation learning framework. Such a framework is essential for embedding structured knowledge into multi-modal Large Language Models effectively, alleviating issues like knowledge misconceptions and multi-modal hallucinations. In this work, we explore the efficacy of models in accurately embedding entities within MMKGs through two pivotal tasks: Multi-modal Knowledge Graph Completion (MKGC) and Multi-modal Entity Alignment (MMEA). Building on this foundation, we propose a novel SNAG method that utilizes a Transformer-based architecture equipped with modality-level noise masking to robustly integrate multi-modal entity features in KGs. By incorporating specific training objectives for both MKGC and MMEA, our approach achieves SOTA performance across a total of ten datasets, demonstrating its versatility. Moreover, SNAG can not only function as a standalone model but also enhance other existing methods, providing stable performance improvements. Code and data are available at https://github.com/zjukg/SNAG.

Modular deep learning is the state-of-the-art solution for lifting the curse of multilinguality, preventing the impact of negative interference and enabling cross-lingual performance in Multilingual Pre-trained Language Models. However, a trade-off of this approach is the reduction in positive transfer learning from closely related languages. In response, we introduce a novel method called language arithmetic, which enables training-free post-processing to address this limitation. Extending the task arithmetic framework, we apply learning via addition to the language adapters, transitioning the framework from a multi-task to a multilingual setup. The effectiveness of the proposed solution is demonstrated on three downstream tasks in a MAD-X-based set of cross-lingual schemes, acting as a post-processing procedure. Language arithmetic consistently improves the baselines with significant gains, especially in the most challenging case of zero-shot application. Our code and models are available at https://github.com/mklimasz/language-arithmetic.

Large language models (LLMs) are playing a pivotal role in deploying strategic use cases across a range of organizations, from large pan-continental companies to emerging startups. The issues and challenges involved in the successful utilization of LLMs can vary significantly depending on the size of the organization. It is important to study and discuss these pertinent issues of LLM adaptation with a focus on the scale of the industrial concerns and brainstorm possible solutions and prospective directions. Such a study has not been prominently featured in the current research literature. In this study, we adopt a threefold strategy: first, we conduct a case study with industry practitioners to formulate the key research questions; second, we examine existing industrial publications to address these questions; and finally, we provide a practical guide for industries to utilize LLMs more efficiently. We release the GitHub repository with the most recent papers in the field.

We introduce a novel Transformer-based method for document segmentation, tailored for practical, real-world applications. This method utilizes overlapping text sequences with a unique position-aware weighting mechanism to enhance segmentation accuracy. Through comprehensive experiments on both public and proprietary datasets, we demonstrate significant improvements, establishing new state-of-the-art standards by achieving up to a 10% increase in segmentation F1 score compared to existing methods. Additionally, we explore the application of our segmentation method in downstream retrieval-augmented question answering tasks, where it improves the quality of generated responses by 5% while achieving up to four times greater efficiency. These results underscore our model’s potential as a robust and scalable solution for real-world text segmentation challenges.

Large language models (LLMs) combined with tool learning have gained impressive results in real-world applications. During tool learning, LLMs may call multiple tools in nested orders, where the latter tool call may take the former response as its input parameters. However, current research on the nested tool learning capabilities is still under-explored, since the existing benchmarks lack relevant data instances. To address this problem, we introduce NesTools to bridge the current gap in comprehensive nested tool learning evaluations. NesTools comprises a novel automatic data generation method to construct large-scale nested tool calls with different nesting structures. With manual review and refinement, the dataset is in high quality and closely aligned with real-world scenarios. Therefore, NesTools can serve as a new benchmark to evaluate the nested tool learning abilities of LLMs. We conduct extensive experiments on 22 LLMs, and provide in-depth analyses with NesTools, which shows that current LLMs still suffer from the complex nested tool learning task.

Large Language Models (LLMs) have shown impressive performance on various benchmarks, yet their ability to engage in deliberate reasoning remains questionable. We present NYT-Connections, a collection of 358 simple word classification puzzles derived from the New York Times Connections game. This benchmark is designed to penalize quick, intuitive “System 1” thinking, isolating fundamental reasoning skills. We evaluated six recent LLMs, a simple machine learning heuristic, and humans across three configurations: single-attempt, multiple attempts without hints, and multiple attempts with contextual hints. Our findings reveal a significant performance gap: even top-performing LLMs like GPT-4 fall short of human performance by nearly 30%. Notably, advanced prompting techniques such as Chain-of-Thought and Self-Consistency show diminishing returns as task difficulty increases. NYT-Connections uniquely combines linguistic isolation, resistance to intuitive shortcuts, and regular updates to mitigate data leakage, offering a novel tool for assessing LLM reasoning capabilities.

The integration of artificial intelligence (AI) in legal judgment prediction (LJP) has the potential to transform the legal landscape, particularly in jurisdictions like India, where a significant backlog of cases burdens the legal system. This paper introduces NyayaAnumana, the largest and most diverse corpus of Indian legal cases compiled for LJP, encompassing a total of 7,02,945 preprocessed cases. NyayaAnumana, which combines the words “Nyaya” and “Anumana” that means “judgment” and “inference” respectively for most major Indian languages, includes a wide range of cases from the Supreme Court, High Courts, Tribunal Courts, District Courts, and Daily Orders and, thus, provides unparalleled diversity and coverage. Our dataset surpasses existing datasets like PredEx and ILDC, offering a comprehensive foundation for advanced AI research in the legal domain. In addition to the dataset, we present INLegalLlama, a domain-specific generative large language model (LLM) tailored to the intricacies of the Indian legal system. It is developed through a two-phase training approach over a base LLaMa model. First, Indian legal documents are injected using continual pretraining. Second, task-specific supervised finetuning is done. This method allows the model to achieve a deeper understanding of legal contexts. Our experiments demonstrate that incorporating diverse court data significantly boosts model accuracy, achieving approximately 90% F1-score in prediction tasks. INLegalLlama not only improves prediction accuracy but also offers comprehensible explanations, addressing the need for explainability in AI-assisted legal decisions.

Torlak is an endangered, low-resource Slavic language with a high degree of areal and inter-speaker variation. In previous work, interviews were performed with Torlak speakers in Serbia, near the Bulgarian border, and the transcripts annotated with lemma and morphosyntactic descriptions at token level. As such token-level annotations facilitate cross-language comparison in the context of the Balkan Sprachbund, where multiple languages influenced Torlak over time, including Serbian and Bulgarian. Here, we aim to improve the prediction of morphosyntactic annotations for this low-resource language using the fine-tuning of large language models, comparing several predictive models. We also further fine-tuned the large language models for scoring the degree of ‘Torlakness’ of a sentence by labeling likely Torlak tokens, to facilitate the documentation of additional Torlak transcribed speech with a high degree of Torlak-style non-standard features compared to standard Serbian. Taken together, we hope that these contributions will help to document this endangered language, and improve digital access for its speakers.

Entity and relation extraction is a conventional task in the field of information extraction. Existing work primarily focuses on detecting specific relations between entities, often constrained to particular fields and lacking general applicability. In response, we propose a novel task: nominal compound relation extraction (NCRE), which concentrates on abstract and broadly applicable relation extraction between noun phrases. This task diverges significantly from traditional entity and relation extraction in two key respects. Firstly, our task involves general nominal compounds rather than named entities, which are longer and encompass a broader scope, presenting significant challenges for extraction. Secondly, relation extraction in NCRE demands an in-depth understanding of context to detect abstract relations. We manually annotate a high-quality Chinese dataset for the NCRE task and develop a model incorporating the rotary position-enhanced word pair (RoWP) detection schema. Experimental results demonstrate the efficiency of our RoWP model over previous baselines, while the suboptimal F1 scores indicate that NCRE remains a challenging task. Our code and data are available at https://github.com/yeecjc/NCRE.

Recent breakthroughs in Large Language Models (LLMs) have led to their adoption across a wide range of tasks, ranging from code generation to machine translation and sentiment analysis, etc. Red teaming/Safety alignment efforts show that fine-tuning models on benign (non-harmful) data could compromise safety. However, it remains unclear to what extent this phenomenon is influenced by different variables, including fine-tuning task, model calibrations, etc. This paper explores the task-wise safety degradation due to fine-tuning on downstream tasks such as summarization, code generation, translation, and classification across various calibration. Our results reveal that: 1) Fine-tuning LLMs for code generation and translation leads to the highest degradation in safety guardrails. 2) LLMs generally have weaker guardrails for translation and classification, with 73-92% of harmful prompts answered, across baseline and other calibrations, falling into one of two concern categories. 3) Current solutions, including guards and safety tuning datasets, lack cross-task robustness. To address these issues, we developed a new multitask safety dataset effectively reducing attack success rates across a range of tasks without compromising the model’s overall helpfulness. Our work underscores the need for generalized alignment measures to ensure safer and more robust models.

This study evaluates the effectiveness of Vision Language Models (VLMs) in representing and utilizing multimodal content for fact-checking. To be more specific, we investigate whether incorporating multimodal content improves performance compared to text-only models and how well VLMs utilize text and image information to enhance misinformation detection. Furthermore we propose a probing classifier based solution using VLMs. Our approach extracts embeddings from the last hidden layer of selected VLMs and inputs them into a neural probing classifier for multi-class veracity classification. Through a series of experiments on two fact-checking datasets, we demonstrate that while multimodality can enhance performance, fusing separate embeddings from text and image encoders yielded superior results compared to using VLM embeddings. Furthermore, the proposed neural classifier significantly outperformed KNN and SVM baselines in leveraging extracted embeddings, highlighting its effectiveness for multimodal fact-checking.

Most research on implicit discourse relation identification has focused on written language, however, it is also crucial to understand these relations in spoken discourse. We introduce a novel method for implicit discourse relation identification across both text and speech, that allows us to extract examples of semantically equivalent pairs of implicit and explicit discourse markers, based on aligning speech+transcripts with subtitles in another language variant. We apply our method to Egyptian Arabic, resulting in a novel high-quality dataset of spoken implicit discourse relations. We present a comprehensive approach to modeling implicit discourse relation classification using audio and text data with a range of different models. We find that text-based models outperform audio-based models, but combining text and audio features can lead to enhanced performance.

Multimodal Aspect-Based Sentiment Analysis (MABSA) aims to extract aspect terms from text-image pairs and identify their sentiments. Previous methods are based on the premise that the image contains the objects referred by the aspects within the text. However, this condition cannot always be met, resulting in a suboptimal performance. In this paper, we propose COnditional Relation based Sentiment Analysis framework (CORSA). Specifically, we design a conditional relation detector (CRD) to mitigate the impact of the unmet conditional image. Moreover, we design a visual object localizer (VOL) to locate the exact condition-related visual regions associated with the aspects. With CRD and VOL, our CORSA framework takes a multi-task form. In addition, to effectively learn CORSA we conduct two types of annotations. One is the conditional relation using a pretrained referring expression comprehension model; the other is the bounding boxes of visual objects by a pretrained object detection model. Experiments on our built C-MABSA dataset show that CORSA consistently outperforms existing methods. The code and data are available at https://github.com/Liuxj-Anya/CORSA.

Even with various regulations in place across countries and social media platforms (Government of India, 2021; European Parliament and Council of the European Union, 2022), digital abusive speech remains a significant issue. One potential approach to address this challenge is automatic text detoxification, a text style transfer (TST) approach that transforms toxic language into a more neutral or non-toxic form. To date, the availability of parallel corpora for the text detoxification task (Logacheva et al., 2022; Atwell et al., 2022; Dementieva et al., 2024a) has proven to be crucial for state-of-the-art approaches. With this work, we extend parallel text detoxification corpus to new languages—German, Chinese, Arabic, Hindi, and Amharic—testing in the extensive multilingual setup TST baselines. Next, we conduct the first of its kind an automated, explainable analysis of the descriptive features of both toxic and non-toxic sentences, diving deeply into the nuances, similarities, and differences of toxicity and detoxification across 9 languages. Finally, based on the obtained insights, we experiment with a novel text detoxification method inspired by the Chain-of-Thoughts reasoning approach, enhancing the prompting process through clustering on relevant descriptive attributes.

Adpositions display a remarkable amount of ambiguity and flexibility in their meanings, and are used in different ways across languages. We conduct a systematic corpus-based cross-linguistic investigation into the lexical semantics of adpositions, utilizing SNACS (Schneider et al., 2018), an annotation framework with data available in several languages. Our investigation encompasses 5 of these languages: Chinese, English, Gujarati, Hindi, and Japanese. We find substantial distributional differences in adposition semantics, even in comparable corpora. We further train classifiers to disambiguate adpositions in each of our languages. Despite the cross-linguistic differences in adpositional usage, sharing annotated data across languages boosts overall disambiguation performance, leading to the highest published scores on this task for all 5 languages.

Multilingual knowledge editing (MKE) aims to simultaneously update factual knowledge across multiple languages within large language models (LLMs). Previous research indicates that the same knowledge across different languages within LLMs exhibits a degree of shareability. However, most existing MKE methods overlook the connections of the same knowledge between different languages, resulting in knowledge conflicts and limited edit performance. To address this issue, we first investigate how LLMs process multilingual factual knowledge and discover that the same factual knowledge in different languages generally activates a shared set of neurons, which we call language-agnostic factual neurons (LAFNs). These neurons represent the same factual knowledge shared across languages and imply the semantic connections among multilingual knowledge. Inspired by this finding, we propose a new MKE method by Locating and Updating Language-Agnostic Factual Neurons (LU-LAFNs) to edit multilingual knowledge simultaneously, which avoids knowledge conflicts and thus improves edit performance. Experimental results on Bi-ZsRE and MzsRE benchmarks demonstrate that our method achieves the best edit performance, indicating the effectiveness and importance of modeling the semantic connections among multilingual knowledge.

Query-product relevance classification is crucial for e-commerce stores like Amazon, ensuring accurate search results that match customer intent. Using a unified multilingual model across multiple languages/marketplaces tends to yield superior outcomes but also presents challenges, especially in maintaining performance across all languages when the model is updated or expanded to include a new one. To tackle this, we examine a multilingual continual learning (CL) framework focused on relevance classification tasks and address the issue of catastrophic forgetting. We propose a novel continual learning approach called attention distillation, which sequentially adds adapters for each new language and incorporates a fusion layer above language-specific adapters. This fusion layer distills attention scores from the previously trained fusion layer, focusing on the older adapters. Additionally, translating a portion of the new language data into older ones supports backward knowledge transfer. Our method reduces trainable parameters by 80%, enhancing computational efficiency and enabling frequent updates, while achieving a 1-3% ROC-AUC improvement over single marketplace baselines and outperforming SOTA CL methods on proprietary and external datasets.

Programming-Community Question Answering (PCQA) aims to tackle issues through generating functional code and guiding descriptions. It involves multiple candidates, with different users having varying preferences for them. Additionally, one may contain outdated APIs. These undoubtedly present a challenge for responsing that meet user preferences. Recently, Reinforcement Learning from Human Feedback demonstrates its ability to precisely control the behavior of large language models (LLMs) to yield human-like responses. However, applying it to LLMs in domain-specific PCQA remains unexplored. In this work, we propose Multi-perspective Preference Alignment for Programming-Community Question Answering to generate user-centric responses, called MupPCQA. It includes three stages: Preference Standardization to control content quality, Preference Integration to consider diverse user tendencies, Preference Timeliness Mitigation to alleviate outdated answers. Extensive experiments on a high-quality, real-world PCQA dataset validate its accuracy and preference. Compared to its base model, MupPCQA shows an improvement of nearly 11% in BLEU, with increases of 20% and 17.5% in BERTScore and CodeBERTScore.

Multimodal sarcasm detection (MSD) is essential for various downstream tasks. Existing MSD methods tend to rely on spurious correlations. These methods often mistakenly prioritize non-essential features yet still make correct predictions, demonstrating poor generalizability beyond training environments. Regarding this phenomenon, this paper undertakes several initiatives. Firstly, we identify two primary causes that lead to the reliance of spurious correlations. Secondly, we address these challenges by proposing a novel method that integrate Multimodal Incongruities via Contrastive Learning (MICL) for multimodal sarcasm detection. Specifically, we first leverage incongruity to drive multi-view learning from three views: token-patch, entity-object, and sentiment. Then, we introduce extensive data augmentation to mitigate the biased learning of the textual modality. Additionally, we construct a test set, SPMSD, which consists potential spurious correlations to evaluate the the model’s generalizability. Experimental results demonstrate the superiority of MICL on benchmark datasets, along with the analyses showcasing MICL’s advancement in mitigating the effect of spurious correlation.