COLING 2025 论文列表

Using tools by Large Language Models (LLMs) is a promising avenue to extend their reach beyond language or conversational settings. The number of tools can scale to thousands as they enable accessing sensory information, fetching updated factual knowledge, or taking actions in the real world. In such settings, in-context learning by providing a short list of relevant tools in the prompt is a viable approach. To retrieve relevant tools, various approaches have been suggested, ranging from simple frequency-based matching to dense embedding-based semantic retrieval. However, such approaches lack the contextual and common-sense understanding required to retrieve the right tools for complex user requests. Rather than increasing the complexity of the retrieval component itself, we propose leveraging LLM understanding to generate a retrieval query. Then, the generated query is embedded and used to find the most relevant tools via a nearest-neighbor search. We investigate three approaches for query generation: zero-shot prompting, supervised fine-tuning on tool descriptions, and alignment learning by iteratively optimizing a reward metric measuring retrieval performance. By conducting extensive experiments on a dataset covering complex and multi-tool scenarios, we show that leveraging LLMs for query generation improves the retrieval for in-domain (seen tools) and out-of-domain (unseen tools) settings.

Relation extraction is a crucial natural language processing task that extracts relational triplets from raw text. Syntactic dependencies information has shown its effectiveness for relation extraction tasks. However, in most existing studies, dependency information is used only for traditional encoder-only-based relation extraction, not for generative sequence-to-sequence (seq2seq)-based relation extraction. In this study, we propose a syntax-aware seq2seq pre-trained model for seq2seq-based relation extraction. The model incorporates dependency information into a seq2seq pre-trained language model by continual pre-training with a seq2seq-based dependency parsing task. Experimental results on two widely used relation extraction benchmark datasets show that dependency parsing pre-training can improve the relation extraction performance.

Reinforcement learning has shown great promise in aligning language models with human preferences in a variety of text generation tasks, including machine translation. For translation tasks, rewards can easily be obtained from quality estimation (QE) models which can generate rewards for unlabeled data. Despite its usefulness, reinforcement learning cannot exploit the gradients with respect to the QE score. We propose QE-EBM, a method of employing quality estimators as trainable loss networks that can directly backpropagate to the NMT model. We examine our method on several low and high resource target languages with English as the source language. QE-EBM outperforms strong baselines such as REINFORCE and proximal policy optimization (PPO) as well as supervised fine-tuning for all target languages, especially low-resource target languages. Most notably, for English-to-Mongolian translation, our method achieves improvements of 2.5 BLEU, 7.1 COMET-KIWI, 5.3 COMET, and 6.4 XCOMET relative to the supervised baseline.

Sign Language Translation (SLT) bridges the communication gap between deaf and hearing individuals by converting sign language videos into spoken language texts. While most SLT research has focused on bilingual translation models, the recent surge in interest has led to the exploration of Multilingual Sign Language Translation (MSLT). However, MSLT presents unique challenges due to the diversity of sign languages across nations. This diversity can lead to cross-linguistic conflicts and hinder translation accuracy. To use the similarity of actions and semantics between sign languages to alleviate conflict, we propose a novel approach that leverages sign language families to improve MSLT performance. Sign languages were clustered into families automatically based on their Language distribution in the MSLT network. We compare the results of our proposed family clustering method with the analysis conducted by sign language linguists and then train dedicated translation models for each family in the many-to-one translation scenario. Our experiments on the SP-10 dataset demonstrate that our approach can achieve a balance between translation accuracy and computational cost by regulating the number of language families.

Automatic fact-checking systems that employ large language models (LLMs) have achieved human-level performance in combating widespread misinformation. However, current LLM-based fact-checking systems fail to reveal the reasoning principles behind their decision-making for the claim verdict. In this work, we propose Correlation-Enhanced Explainable Fact-Checking (CorXFact), an LLM-based fact-checking system that simulates the reasoning principle of human fact-checkers for evidence-based claim verification: assessing and weighing the correlations between the claim and each piece of evidence. Following this principle, CorXFact enables efficient claim verification and transparent explanation generation. Furthermore, we contribute the CorFEVER test set to comprehensively evaluate the CorXFact system in claim-evidence correlation identification and claim verification in both closed-domain and real-world fact-checking scenarios. Experimental results show that our proposed CorXFact significantly outperforms four strong fact-checking baselines in claim authenticity prediction and verdict explanation.

Comprehensive error annotation is essential for developing effective Grammatical Error Correction (GEC) systems and delivering meaningful feedback to learners. This paper introduces improvements to automatic grammatical error annotation for Chinese. Our refined framework addresses language-specific challenges that cause common spelling errors in Chinese, including pronunciation similarity, visual shape similarity, specialized participles, and word ordering. In a case study, we demonstrated our system’s ability to provide detailed feedback on 12-16% of all errors by identifying them under our new error typology, specific enough to uncover subtle differences in error patterns between L1 and L2 writings. In addition to improving automated feedback for writers, this work also highlights the value of incorporating language-specific features in NLP systems.

The opinions of the U.S. Supreme Court (SCOTUS) are known for their extensive length, complex legal language, and lack of titled sections, which pose significant challenges for accessibility and comprehension. This paper defines the task of automatic section titling by proposing both generic and specific headings for each section. Given the scarcity of sections with headings in SCOTUS, we study the possibility of using data from lower courts for training models. A dataset of sections with generic or specific headings covering three courts (SCOTUS and two lower courts) was compiled. A supplementary SCOTUS set was manually annotated with these two types of titles. In order to establish a benchmark, we provide the performance of different systems trained for each subtask: For generic heading prediction, we compare the performance of fine-tuning non-contextual, general and domain-oriented pretrained language models. Transformer-based sequence-to-sequence models are considered for specific heading generation. Our results show that a fine-tuned LegalBERT can achieve a F1 score of about 0.90 % in predicting generic headings. They also show that BART and T5 have similar performance in generating specific headings and that, although this performance is good, there is still room for improvement. In addition, we provide a human assessment to support the generation experiment and show a quasi-linear correlation between human degrees of agreement and the results of conventional measures such as ROUGE and BERTScore.

Current sign language translation (SLT) approaches often rely on gloss-based supervision with Connectionist Temporal Classification (CTC), limiting their ability to handle non-monotonic alignments between sign language video and spoken text. In this work, we propose a novel method combining joint CTC/Attention and transfer learning. The joint CTC/Attention introduces hierarchical encoding and integrates CTC with the attention mechanism during decoding, effectively managing both monotonic and non-monotonic alignments. Meanwhile, transfer learning helps bridge the modality gap between vision and language in SLT. Experimental results on two widely adopted benchmarks, RWTH-PHOENIX-Weather 2014 T and CSL-Daily, show that our method achieves results comparable to state-of-the-art and outperforms the pure-attention baseline. Additionally, this work opens a new door for future research into gloss-free SLT using text-based CTC alignment.

The Mixture-of-Experts (MoE) architecture has demonstrated significant potential in both large-scale pre-training and instruction tuning by offering increased parameter capacity without additional inference costs. However, developing MoE models faces challenges including training instability and the need for substantial high-quality training data. While efficient methodologies like sparse upcycling exist, they often lead to performance degradation in instruction tuning scenarios. We introduce representation-based sparse upcycling, a straightforward yet effective technique for converting dense language models into sparsely activated ones while maintaining similar computational costs. Unlike conventional sparse upcycling, our approach leverages intermediate representations from language models to initialize router weights. This strategy addresses the mismatch between randomly initialized and well-trained parameters while providing prior knowledge to guide expert specialization during training. Extensive experiments across diverse benchmarks demonstrate significant improvements in both model capabilities and routing consistency compared to existing approaches.

Despite recent progress in Speech Translation (ST) research, the challenges posed by inherent speech phenomena that distinguish transcribed speech from written text are not well addressed. The informal and erroneous nature of spontaneous speech is inadequately represented in the typical parallel text available for building translation models. We propose to address these issues through a text rewrite approach that aims to transform transcribed speech into a cleaner style more in line with the expectations of translation models built from written text. Moreover, the advantages of the rewrite model can be effectively distilled into a standalone translation model. Experiments on several benchmarks, using both publicly available and in-house translation models, demonstrate that adding a rewrite model to a traditional ST pipeline is a cost-effect way to address a variety of speech irregularities and improve speech translation quality for multiple language directions and domains.

Detecting euphemisms is essential for content security on various social media platforms, but existing methods designed for detecting euphemisms are ineffective in impromptu euphemisms. In this work, we make a first attempt to an exploration of impromptu euphemism detection and introduce the Impromptu Cybercrime Euphemisms Detection (ICED) dataset. Moreover, we propose a detection framework tailored to this problem, which employs context augmentation modeling and multi-round iterative training. Our detection framework mainly consists of a coarse-grained and a fine-grained classification model. The coarse-grained classification model removes most of the harmless content in the corpus to be detected. The fine-grained model, impromptu euphemisms detector, integrates context augmentation and multi-round iterations training to better predicts the actual meaning of a masked token. In addition, we leverage ChatGPT to evaluate the mode’s capability. Experimental results demonstrate that our approach achieves a remarkable 76-fold improvement compared to the previous state-of-the-art euphemism detector.

We present ImPoster, a novel algorithm for generating a target image of a ‘source’ subject performing a ‘driving’ action. The inputs to our algorithm are a single pair of a source image with the subject that we wish to edit and a driving image with a subject of an arbitrary class performing the driving action, along with the text descriptions of the two images. Our approach is completely unsupervised and does not require any access to additional annotations like keypoints or pose. Our approach builds on a pretrained text-to-image latent diffusion model and learns the characteristics of the source and the driving image by finetuning the diffusion model for a small number of iterations. At inference time, ImPoster performs step-wise text prompting i.e. it denoises by first moving in the direction of the image manifold corresponding to the driving image followed by the direction of the image manifold corresponding to the text description of the desired target image. We propose a novel diffusion guidance formulation, image frequency guidance, to steer the generation towards the manifold of the source subject and the driving action at every step of the inference denoising. Our frequency guidance formulations are derived from the frequency domain properties of images. We extensively evaluate ImPoster on a diverse set of source-driving image pairs to demonstrate improvements over baselines. To the best of our knowledge, ImPoster is the first approach towards achieving both subject-driven as well as action-driven image personalization.

Nigerian Pidgin (NP) is an English-based creole language spoken by nearly 100 million people across Nigeria, and is still low-resource in NLP. In particular, there are currently no available discourse parsing tools, which, if available, would have the potential to improve various downstream tasks. Our research focuses on implicit discourse relation classification (IDRC) for NP, a task which, even in English, is not easily solved by prompting LLMs, but requires supervised training. % With this in mind, we have developed a framework for the task, which could also be used by researchers for other English-lexified languages. We systematically compare different approaches to the low resource IDRC task: in one approach, we use English IDRC tools directly on the NP text as well as on their English translations (followed by a back-projection of labels). In another approach, we create a synthetic discourse corpus for NP, in which we automatically translate the English discourse-annotated corpus PDTB to NP, project PDTB labels, and then train an NP IDR classifier. The latter approach of training a “native” NP classifier outperforms our baseline by 13.27% and 33.98% in f1 score for 4-way and 11-way classification, respectively.

With the success of 2D diffusion models, 2D AIGC content has already transformed our lives. Recently, this success has been extended to 3D AIGC, with state-of-the-art methods generating textured 3D models from single images or text. However, we argue that current 3D AIGC methods still don’t fully unleash human creativity. We often imagine 3D content made from multimodal inputs, such as what it would look like if my pet bunny were eating a doughnut on the table. In this paper, we explore a novel 3D AIGC approach: generating 3D content from IDEAs. An IDEA is a multimodal input composed of text, image, and 3D models. To our knowledge, this challenging and exciting 3D AIGC setting has not been studied before. We propose the new framework Idea23D, which combines three agents based on large multimodal models (LMMs) and existing algorithmic tools. These three LMM-based agents are tasked with prompt generation, model selection, and feedback reflection. They collaborate and critique each other in a fully automated loop, without human intervention. The framework then generates a text prompt to create 3D models that align closely with the input IDEAs. We demonstrate impressive 3D AIGC results that surpass previous methods. To comprehensively assess the 3D AIGC capabilities of Idea23D, we introduce the Eval3DAIGC-198 dataset, containing 198 multimodal inputs for 3D generation tasks. This dataset evaluates the alignment between generated 3D content and input IDEAs. Our user study and quantitative results show that Idea23D significantly improves the success rate and accuracy of 3D generation, with excellent compatibility across various LMM, Text-to-Image, and Image-to-3D models. Code and dataset are available at https://idea23d.github.io/.

The current best practice to measure the performance of base Large Language Models is to establish a multi-task benchmark that covers a range of capabilities of interest. Currently, however, such benchmarks are only available in a few high-resource languages. To address this situation, we present IberoBench, a multilingual, multi-task benchmark for Iberian languages (i.e., Basque, Catalan, Galician, European Spanish and European Portuguese) built on the LM Evaluation Harness framework. The benchmark consists of 62 tasks divided into 179 subtasks. We evaluate 33 existing LLMs on IberoBench on 0- and 5-shot settings. We also explore the issues we encounter when working with the Harness and our approach to solving them to ensure high-quality evaluation.

Inspired by human cognitive behavior, we introduce visual modality to enhance the performance of pure text-based question-answering tasks with the development of multimodal models. However, obtaining corresponding images through manual annotation often entails high costs. Faced with this challenge, an intuitive strategy is to use search engines or use web scraping techniques to automatically obtain relevant image information. However, the images obtained by this strategy may be of low quality and may not match the context of the original task, which could fail to improve or even decrease performance on downstream tasks. In this paper, we propose a novel framework named “ITERATE”, aimed at retrieving and optimizing the quality of images to improve the alignment between text and images. Inspired by evolutionary algorithms in reinforcement learning and driven by the synergy of large language models (LLMs) and multimodal models, ITERATE employs a series of strategic actions such as filtering, optimizing, and retrieving to acquire higher quality images, and repeats this process over multiple generations to enhance the quality of the entire image cluster. Our experimental results on the ScienceQA, ARC-Easy, and OpenDataEval datasets also verify the effectiveness of our method, showing improvements of 3.5%, 5%, and 7%, respectively.

In this paper, we present the IRUEX dataset, a novel multiple-choice educational resource specifically designed to evaluate the performance of Large Language Models (LLMs) across seven distinct categories. The dataset contains 868 Iran university entrance exam questions (Konkour) and 36,485 additional questions. Each additional question is accompanied by detailed solutions, and the dataset also includes relevant high school textbooks, providing comprehensive study material. A key feature of IRUEX is its focus on underrepresented languages, particularly assessing problem-solving skills, language proficiency, and reasoning. Our evaluation shows that GPT-4o outperforms the other LLMs tested on the IRUEX dataset. Techniques such as few-shot learning and retrieval-augmented generation (RAG) display varied effects across different categories, highlighting their unique strengths in specific areas. Additionally, a comprehensive user study classifies the errors made by LLMs into ten problem-solving ability categories. The analysis highlights that calculations and linguistic knowledge, particularly in low-resource languages, remain significant weaknesses in current LLMs. IRUEX has the potential to serve as a benchmark for evaluating the reasoning capabilities of LLMs in non-English settings, providing a foundation for improving their performance in diverse languages and contexts

Large-scale Vision-Language Models (LVLMs) process both images and text, excelling in multimodal tasks such as image captioning and description generation. However, while these models excel at generating factual content, their ability to generate and evaluate texts reflecting perspectives on the same image, depending on the context, has not been sufficiently explored. To address this, we propose IRR: Image Review Rank, a novel evaluation framework designed to assess critic review texts from multiple perspectives. IRR evaluates LVLMs by measuring how closely their judgments align with human interpretations. We validate it using a dataset of images from 15 categories, each with five critic review texts and annotated rankings in both English and Japanese, totaling over 2,000 data instances. Our results indicate that, although LVLMs exhibited consistent performance across languages, their correlation with human annotations was insufficient, highlighting the need for further advancements. These findings highlight the limitations of current evaluation methods and the need for approaches that better capture human reasoning in Vision & Language tasks.

In-Context Learning (ICL) is a critical capability of Large Language Models (LLMs) as it empowers them to comprehend and reason across interconnected inputs. Evaluating the ICL ability of LLMs can enhance their utilization and deepen our understanding of how this ability is acquired at the training stage. However, existing evaluation frameworks primarily focus on language abilities and knowledge, often overlooking the assessment of ICL ability. In this work, we introduce the ICLEval benchmark to evaluate the ICL abilities of LLMs, which encompasses two key sub-abilities: exact copying and rule learning. Through the ICLEval benchmark, we demonstrate that ICL ability is universally present in different LLMs, and model size is not the sole determinant of ICL efficacy. Surprisingly, we observe that ICL abilities, particularly copying, develop early in the pretraining process and stabilize afterward.