COLING 2025 Papers

Large Language Models (LLMs) have rapidly advanced, with domain-specific expert models emerging to handle specialized tasks across various fields. However, the predominant focus on English-centric models demands extensive data, making it challenging to develop comparable models for middle and low-resource languages. To address this limitation, we introduce Migrate, a novel method that leverages open-source static embedding models and up to 3 million tokens of code-switching data to facilitate the seamless transfer of embeddings to target languages. Migrate enables effective cross-lingual adaptation without requiring large-scale domain-specific corpora in the target language, promoting the accessibility of expert LLMs to a diverse range of linguistic communities. Our experimental results demonstrate that Migrate significantly enhances model performance in target languages, outperforming baseline and existing cross-lingual transfer methods. This approach provides a practical and efficient solution for extending the capabilities of domain-specific expert models.

Decoder-only Large Language Models (LLMs) have demonstrated exceptional performance in language generation, exhibiting broad capabilities across various tasks. However, the application to label-sensitive language understanding tasks remains challenging due to the limitations of their autoregressive architecture, which restricts the sharing of token information within a sentence. In this paper, we address the Multi-Intent Detection (MID) task and introduce MIDLM, a bidirectional LLM framework that incorporates intent number detection and multi-intent selection. This framework allows autoregressive LLMs to leverage bidirectional information awareness through post-training, eliminating the need for training the models from scratch. Comprehensive evaluations across 8 datasets show that MIDLM consistently outperforms both existing vanilla models and pretrained baselines, demonstrating its superior performance in the MID task.

Supervised fine-tuning (SFT) is widely adopted for tailoring large language models (LLMs) to specific downstream tasks. However, the substantial computational demands of LLMs hinder iterative exploration of fine-tuning datasets and accurate evaluation of individual sample importance. To address this challenge, we introduce Meta-LoRA, a memory-efficient method for automatic sample reweighting. Meta-LoRA learns to reweight fine-tuning samples by minimizing the loss on a small, high-quality validation set through an end-to-end bi-level optimization framework based on meta-learning. To reduce memory usage associated with computing second derivatives, we approximate the bi-level optimization using gradient similarity between training and validation datasets, replacing bi-dimensional gradient similarity with the product of one-dimensional activation states and their corresponding gradients. Further memory optimization is achieved by refining gradient computations, selectively applying them to the low-rank layers of LoRA, which results in as little as 4% additional memory usage. Comprehensive evaluations across benchmark datasets in mathematics, coding, and medical domains demonstrate Meta-LoRA’s superior efficacy and efficiency. The source code is available at https://github.com/liweicheng-ai/meta-lora.

We introduce MESAQA, a novel dataset focusing on multi-span contextual understanding question answering (QA).Unlike traditional single-span QA systems, questions in our dataset consider information from multiple spans within the context document. MESAQA supports evidence-grounded QA, demanding the model’s capability of answer generation and multi-evidence identification. Our automated dataset creation method leverages the MASH-QA dataset and large language models (LLMs) to ensure that each Q/A pair requires considering all selected spans. Experimental results show that current models struggle with multi-span contextual QA, underscoring the need for new approaches. Our dataset sets a benchmark for this emerging QA paradigm, promoting research in complex information retrieval and synthesis.

Moralities, emotions, and events are complex aspects of human cognition, which are often treated separately since capturing their combined effects is challenging, especially due to the lack of annotated data. Leveraging their interrelations hence becomes crucial for advancing the understanding of human moral behaviors. In this work, we propose ME2-BERT, the first holistic framework for fine-tuning a pre-trained language model like BERT to the task of moral foundation prediction. ME2-BERT integrates events and emotions for learning domain-invariant morality-relevant text representations. Our extensive experiments show that ME2-BERT outperforms existing state-of-the-art methods for moral foundation prediction, with an average increase up to 35% in the out-of-domain scenario.

With the extensive use of large language models, automatically generating QA datasets for domain-specific fine-tuning has become crucial. However, considering the multifaceted demands for readability, diversity, and comprehensiveness of QA data, current methodologies fall short in producing high-quality QA datasets. Moreover, the dependence of existing evaluation metrics on ground truth labels further exacerbates the challenges associated with the selection of QA data. In this paper, we introduce a novel method for QA data generation, denoted as MDPO. We proposes a set of unsupervised evaluation metrics for QA data, enabling multidimensional assessment based on the relationships among context,question and answer. Furthermore, leveraging these metrics, we implement a customized direct preference optimization process that guides large language models to produce high-quality and domain-specific QA pairs. Empirical results on public datasets indicate that MDPO’s performance substantially surpasses that of state-of-the-art methods.

Recent advancements in large language models (LLMs) have enabled in-context learning (ICL)-based methods that significantly outperform fine-tuning approaches for text-to-SQL tasks. However, their performance is still considerably lower than that of human experts on benchmarks that include complex schemas and queries, such as BIRD. This study considers the sensitivity of LLMs to the prompts and introduces a novel approach that leverages multiple prompts to explore a broader search space for possible answers and effectively aggregate them. Specifically, we robustly refine the database schema through schema linking using multiple prompts. Thereafter, we generate various candidate SQL queries based on the refined schema and diverse prompts. Finally, the candidate queries are filtered based on their confidence scores, and the optimal query is obtained through a multiple-choice selection that is presented to the LLM. When evaluated on the BIRD and Spider benchmarks, the proposed method achieved execution accuracies of 65.5% and 89.6%, respectively, significantly outperforming previous ICL-based methods.

Retrieval Augmented Generation (RAG) has proven to be highly effective in boosting the generative performance of language model in knowledge-intensive tasks. However, existing RAG framework either indiscriminately perform retrieval or rely on rigid single-label classifiers to select retrieval methods, leading to inefficiencies and suboptimal performance across queries of varying complexity. To address these challenges, we propose a reinforcement learning-based framework that dynamically selects the most suitable retrieval strategy based on query complexity. To address these challenges, we propose a reinforcement learning-based framework that dynamically selects the most suitable retrieval strategy based on query complexity. Our approach leverages a multi-armed bandit algorithm, which treats each retrieval method as a distinct “arm” and adapts the selection process by balancing exploration and exploitation. Additionally, we introduce a dynamic reward function that balances accuracy and efficiency, penalizing methods that require more retrieval steps, even if they lead to a correct result. Our method achieves new state of the art results on multiple single-hop and multi-hop datasets while reducing retrieval costs. Our code are available at https://github.com/FUTUREEEEEE/MBA.

With the quick advancement in text generation ability of Large Language Mode(LLM), concerns about the misuse of machine-generated content have grown, raising potential violations of legal and ethical standards. Some existing studies concentrate on detecting machine-generated text in open-source models using in-model features, but their performance on closed-source large models is limited. This limitation occurs because, in the closed-source model detection, the only reference that can be obtained is the texts, which may differ significantly due to random sampling. In this paper, we demonstrate that texts generated by the same model can align both semantically and statistically under similar prompts, facilitating effective detection and traceability. Specifically, we fine-tune a BERT encoder through contrastive learning to achieve semantic alignment in randomly generated texts from the same model. Then, we propose a method called Machine-Generated Text Detection with Rewritten Texts, which designed several prompt refactoring methods and used them to request rewritten text from LLMs. Semantic and statistical relationships between rewritten and original texts provide a basis for detection and traceability. Finally, we expanded the text dataset with multi-parameter random sampling and verified the performance of MAGRET on three text-generated datasets. Experimental results show that previous methods struggle with closed-source model detection, while our approach significantly outperforms baseline methods in this regard. It also shows MagRet’s stable performance in detection and tracing tasks across various randomly sampled texts.

Recent LLM-based Text-to-SQL methods usually suffer from significant performance degradation on “huge” databases and complex user questions that require multi-step reasoning. Moreover, most existing methods neglect the crucial significance of LLMs utilizing external tools and model collaboration. To address these challenges, we introduce MAC-SQL, a novel LLM-based multi-agent collaborative framework. Our framework comprises a core decomposer agent for Text-to-SQL generation with few-shot chain-of-thought reasoning, accompanied by two auxiliary agents that utilize external tools or models to acquire smaller sub-databases and refine erroneous SQL queries. The decomposer agent collaborates with auxiliary agents, which are activated as needed and can be expanded to accommodate new features or tools for effective Text-to-SQL parsing. In our framework, We initially leverage GPT-4 as the strong backbone LLM for all agent tasks to determine the upper bound of our framework. We then fine-tune an open-sourced instruction-followed model, SQL-Llama, by leveraging Code Llama 7B, to accomplish all tasks as GPT-4 does. Experiments show that SQL-Llama achieves a comparable execution accuracy of 43.94, compared to the baseline accuracy of 46.35 for vanilla GPT-4. At the time of writing, MAC-SQL+GPT-4 achieves an execution accuracy of 59.59 when evaluated on the BIRD benchmark, establishing a new state-of-the-art (SOTA) on its holdout test set.

Sentence embedding models play a key role in various Natural Language Processing tasks, such as in Topic Modeling, Document Clustering and Recommendation Systems. However, these models rely heavily on parallel data, which can be scarce for many low-resource languages, including Luxembourgish. This scarcity results in suboptimal performance of monolingual and cross-lingual sentence embedding models for these languages. To address this issue, we compile a relatively small but high-quality human-generated cross-lingual parallel dataset to train LuxEmbedder, an enhanced sentence embedding model for Luxembourgish with strong cross-lingual capabilities. Additionally, we present evidence suggesting that including low-resource languages in parallel training datasets can be more advantageous for other low-resource languages than relying solely on high-resource language pairs. Furthermore, recognizing the lack of sentence embedding benchmarks for low-resource languages, we create a paraphrase detection benchmark specifically for Luxembourgish, aiming to partially fill this gap and promote further research.

Retriever-Augmented Generation (RAG) systems have become pivotal in enhancing the capabilities of language models by incorporating external knowledge retrieval mechanisms. However, a significant challenge in deploying these systems in industry applications is the detection and mitigation of hallucinations - instances where the model generates information that is not grounded in the retrieved context. Addressing this issue is crucial for ensuring the reliability and accuracy of responses generated by large language models (LLMs) in industry settings. Current hallucination detection techniques fail to deliver accuracy, low latency, and low cost simultaneously. We introduce Luna: a DeBERTA-large encoder, fine-tuned for hallucination detection in RAG settings. We demonstrate that Luna outperforms GPT-3.5 and commercial evaluation frameworks on the hallucination detection task, with 97% and 91% reduction in cost and latency, respectively. Luna’s generalization capacity across multiple industry verticals and out-of-domain data makes it a strong candidate for guardrailing industry LLM applications.

Although large language models have significantly advanced natural language generation, their potential in low-resource machine translation has not yet been fully explored, especially for languages that translation models have not been trained on. In this study, we provide a detailed demonstration of how to efficiently expand low-resource languages for large language models and significantly enhance the model’s translation ability, using Uyghur as an example. The process involves four stages: collecting and pre-processing monolingual data, conducting continuous pre-training with extensive monolingual data, fine-tuning with less parallel corpora using translation supervision, and proposing a direct preference optimization based on translation self-evolution (DPOSE) on this basis. Extensive experiments have shown that our strategy effectively expands the low-resource languages supported by large language models and significantly enhances the model’s translation ability in Uyghur with less parallel data. Our research provides detailed insights for expanding other low-resource languages into large language models.

In recent years, text classification methods based on neural networks and pre-trained models have gained increasing attention and demonstrated excellent performance. However, these methods still have some limitations in practical applications: (1) They typically focus only on the matching similarity between sentences. However, there exists implicit high-value information both within sentences of the same class and across different classes, which is very crucial for classification tasks. (2) Existing methods such as pre-trained language models and graph-based approaches often consume substantial memory for training and text-graph construction. (3) Although some low-resource methods can achieve good performance, they often suffer from excessively long processing times. To address these challenges, we propose a low-resource and fast text classification model called LFTC. Our approach begins by constructing a compressor list for each class to fully mine the regularity information within intra-class data. We then remove redundant information irrelevant to the target classification to reduce processing time. Finally, we compute the similarity distance between text pairs for classification. We evaluate LFTC on 9 publicly available benchmark datasets, and the results demonstrate significant improvements in performance and processing time, especially under limited computational and data resources, highlighting its superior advantages.

Reduplication and repetition, though similar in form, serve distinct linguistic purposes. Reduplication is a deliberate morphological process used to express grammatical, semantic, or pragmatic nuances, while repetition is often unintentional and indicative of disfluency. This paper presents the first large-scale study of reduplication and repetition in speech using computational linguistics. We introduce IndicRedRep, a new publicly available dataset containing Hindi, Telugu, and Marathi text annotated with reduplication and repetition at the word level. We evaluate transformer-based models for multi-class reduplication and repetition token classification, utilizing the Reparandum-Interregnum-Repair structure to distinguish between the two phenomena. Our models achieve macro F1 scores of up to 85.62% in Hindi, 83.95% in Telugu, and 84.82% in Marathi for reduplication-repetition classification.

Traditionally, argument mining research has approached the task of automatic identification of argument structures by using existing definitions of what constitutes an argument, while leaving the equally important matter of what does not qualify as an argument unaddressed. With the ability to distinguish between what is and what is not a natural language argument being at the core of argument mining as a field, it is interesting that no previous work has explored approaches to effectively select non-related propositions (i.e., propositions that are not connected through an argumentative relation, such as support or attack) that improve the data for learning argument mining tasks better. In this paper, we address the question of how to effectively sample non-related propositions from six different argument mining corpora belonging to different domains and encompassing both monologue and dialogue forms of argumentation. To that end, in addition to considering undersampling baselines from previous work, we propose three new sampling strategies relying on context (i.e., short/long) and the semantic similarity between propositions. Our results indicate that using more informed sampling strategies improves the performance, not only when evaluating models on their respective test splits, but also in the case of cross-domain evaluation.

Recently, Large Vision-Language Models (LVLMs) have demonstrated impressive capabilities in multi-modal context comprehension. However, they still suffer from hallucination problems referring to generating inconsistent outputs with the image content. To mitigate hallucinations, previous studies mainly focus on retraining LVLMs with custom datasets. Although effective, they inherently come with additional computational costs. In this paper, we propose a training-free framework, MVP, that aims to reduce hallucinations by making the most of the innate capabilities of the LVLMs via Multi-View Multi-Path Reasoning. Specifically, we first devise a multi-view information-seeking strategy to thoroughly perceive the comprehensive information in the image, which enriches the general global information captured by the original vision encoder in LVLMs. Furthermore, during the answer decoding, we propose multi-path reasoning for each information view to quantify and aggregate the certainty scores for each potential answer among multiple decoding paths and finally decide the output answer. By fully grasping the information in the image and carefully considering the certainty of the potential answers when decoding, our MVP can effectively reduce hallucinations in LVLMs. The extensive experiments verify that our proposed MVP significantly mitigates the hallucination problem across four well-known LVLMs.

We introduce Loki, an open-source tool designed to address the growing problem of misinformation. Loki adopts a human-centered approach, striking a balance between the quality of fact-checking and the cost of human involvement. It decomposes the fact-checking task into a five-step pipeline: breaking down long texts into individual claims, assessing their check-worthiness, generating queries, retrieving evidence, and verifying the claims. Instead of fully automating the claim verification process, provides essential information at each step to assist human judgment, especially for general users such as journalists and content moderators. Moreover, it has been optimized for latency, robustness, and cost efficiency at a commercially usable level. Loki is released under an MIT license and is available on GitHub. We also provide a video presenting the system and its capabilities.

Logical reasoning is a crucial factor in machine reading comprehension tasks (MRC). Existing methods suffer from the balance between semantic and explicit logical relation representations, in which some emphasize contextual semantics, while others pay more attention to explicit logical features. Additionally, previous methods utilize graph convolutional networks (GCN) for node updates, still exhibiting some shortcomings. To address these challenges, in this paper, we propose a logical reasoning method with contrastive learning and lightweight graph networks (LogiGraph). Our method focuses on the lightweight aspect of the GCN, which greatly improves the shortcomings of the GCN, and employs conjunction and punctuation marks as two types of edges to construct a dual graph. Besides, we combine contrastive learning with graph reasoning, which changes the logical expression’s content as the negative sample of the original context, enabling the model to capture negative logical relationships and improving generalization ability. We conduct extensive experiments on two public datasets, ReClor and LogiQA. Experimental results demonstrate that LogiGraph can achieve state-of-the-art performance on both datasets.