COLING 2025 论文列表

Large language models (LLMs) have made great progress in classification and text generation tasks. However, they are mainly trained on English data and often struggle with low-resource languages. In this study, we explore adding a new language, i.e., Persian, to Llama (a model with a limited understanding of Persian) using parameter-efficient fine-tuning. We employ a multi-stage approach involving pretraining on monolingual Persian data, aligning representations through bilingual pretraining and instruction datasets, and instruction-tuning with task-specific datasets. We evaluate the model’s performance at each stage on generation and classification tasks. Our findings suggest that incorporating the Persian language, through bilingual data alignment, can enhance classification accuracy for Persian tasks, with no adverse impact and sometimes even improvements on English tasks. Additionally, the results highlight the model’s initial strength as a critical factor when working with limited training data, with cross-lingual alignment offering minimal benefits for the low-resource language. Knowledge transfer from English to Persian has a marginal effect, primarily benefiting simple classification tasks.

Previous work adopts large language models (LLMs) as evaluators to evaluate natural language process (NLP) tasks. However, certain shortcomings, e.g., fairness, scope, and accuracy, persist for current LLM evaluators. To analyze whether LLMs can serve as reliable alternatives to humans, we examine the fine-grained alignment between LLM evaluators and human annotators, particularly in understanding the target evaluation tasks and conducting evaluations that meet diverse criteria. This paper explores both conventional tasks (e.g., story generation) and alignment tasks (e.g., math reasoning), each with different evaluation criteria. Our analysis shows that 1) LLM evaluators can generate unnecessary criteria or omit crucial criteria, resulting in a slight deviation from the experts. 2) LLM evaluators excel in general criteria, such as fluency, but face challenges with complex criteria, such as numerical reasoning. We also find that LLM-pre-drafting before human evaluation can help reduce the impact of human subjectivity and minimize annotation outliers in pure human evaluation, leading to more objective evaluation. All resources are available at https://github.com/qtli/CoEval.

Recent improvements in the quality of the generations by large language models have spurred research into identifying machine-generated text. Such work often presents high-performing detectors. However, humans and machines can produce text in different styles and domains, yet the the performance impact of such on machine generated text detection systems remains unclear. In this paper, we audit the classification performance for detecting machine-generated text by evaluating on texts with varying writing styles. We find that classifiers are highly sensitive to stylistic changes and differences in text complexity, and in some cases degrade entirely to random classifiers. We further find that detection systems are particularly susceptible to misclassify easy-to-read texts while they have high performance for complex texts, leading to concerns about the reliability of detection systems. We recommend that future work attends to stylistic factors and reading difficulty levels of human-written and machine-generated text.

Multi-modal entity alignment aims to identify equivalent entities between two different multi-modal knowledge graphs, which consist of structural triples and images associated with entities. Unfortunately, prior works fuse the multi-modal knowledge of all entities only via solely one single fusion strategy. Therefore, the impact of the fusion strategy on individual entities could be largely ignored. To solve this challenge, we propose AMF2SEA, an adaptive multi-modal feature fusion strategy for entity alignment, which dynamically selects the optimal entity-level feature fusion strategy. Additionally, we build a new dataset based on DBP15K, which includes a full set of entity images from multiple inconsistent web sources, making it more representative of the real world. Experimental results demonstrate that our model achieves state-of-the-art (SOTA) performance compared to models using the same modality on DBP15K and its variants with richer image sources and styles. Our code and data are available at https://github.com/ChenxiaoLiJoe/AMFFSEA.

This paper investigates the extent to which LLMs align with humans when personality shifts are associated with language changes. Based on three experiments, that focus on GPT-4o and the Eysenck Personality Questionnaire-Revised (EPQR-A), our initial results reveal a weak yet significant variation in GPT-4o’s personality across languages, indicating that some stem from a language-switching effect rather than translation. Further analysis across five English-speaking countries shows that GPT-4o, leveraging stereotypes, reflects distinct country-specific personality traits.

With the emergence of social media and e-commerce platforms, accurate user profiling has become increasingly vital for recommendation systems and personalized services. Recent studies have focused on generating detailed user profiles by extracting various aspects of user attributes from textual reviews. Nevertheless, these investigations have not fully exploited the potential of the abundant multimodal data at hand. In this study, we propose a novel task called multimodal user profiling. This task emphasizes the utilization of both review texts and their accompanying images to create comprehensive user profiles. By integrating textual and visual data, we leverage their complementary strengths, enabling the generation of more holistic user representations. Additionally, we explore a unified joint training framework with various multimodal training strategies that incorporate users’ historical review texts and images for user profile generation. Our experimental results underscore the significance of multimodal data in enhancing user profile generation and demonstrate the effectiveness of the proposed unified joint training approach.

In this paper, we investigate Extractive Question Answering (EQA) with Large Language Models (LLMs) under domain drift, i.e., can LLMs generalize to domains that require specific knowledge such as medicine and law in a zero-shot fashion without additional in-domain training? To this end, we devise a series of experiments to explain the performance gap empirically. Our findings suggest that: (a) LLMs struggle with dataset demands of closed do- mains such as retrieving long answer spans; (b) Certain LLMs, despite showing strong overall performance, display weaknesses in meeting basic requirements as discriminating between domain-specific senses of words which we link to pre-processing decisions; (c) Scaling model parameters is not always effective for cross-domain generalization; and (d) Closed-domain datasets are quantitatively much different than open-domain EQA datasets and current LLMs struggle to deal with them. Our findings point out important directions for improving existing LLMs.

Detailed descriptions of human motion are crucial for effective fitness training, which highlights the importance of research in fine-grained human motion video captioning. Existing video captioning models often fail to capture the nuanced semantics of videos, resulting in the generated descriptions that are coarse and lack details, especially when depicting human motions. To benchmark the Body Fitness Training scenario, in this paper, we construct a fine-grained human motion video captioning dataset named BoFiT and design a state-of-the-art baseline model named BoFiT-Gen (Body Fitness Training Text Generation). BoFiT-Gen makes use of computer vision techniques to extract angular representations of human motions from videos and LLMs to generate fine-grained descriptions of human motions via prompting. Results show that BoFiT-Gen outperforms previous methods on comprehensive metrics. We aim for this dataset to serve as a useful evaluation set for visio-linguistic models and drive further progress in this field. Our dataset is released at https://github.com/colmon46/bofit.

Previous studies of face-to-face (f2f) communication have suggested that speakers rely heavily on a variety of multi-modal cues to make real-time predictions about upcoming words in rapid turn-taking. To understand how computer-mediated communication (CMC) differs from f2f communication in terms of the prediction mechanism, this study assessed how the loss of multi-modal cues would affect word predictability in turn-taking. Participants watched videos, listened to audio, or read a transcript of f2f conversations. Across these three conditions, they predicted the same set of omitted words with different levels of predictability and semantic relatedness to other words in the context. Results showed that words of higher predictability were more accurately predicted regardless of CMC types. Higher response accuracy but longer response time were observed in conditions with richer cues, and for participants with more positive and less negative self-emotions. Meanwhile, semantic relatedness did not affect predictability. These results confirmed the key role of prediction in language processing and conversation smoothness, especially its importance in CMC.

Large language models (LLMs) have shown remarkable performances across a wide range of tasks. However, the mechanisms by which these models encode tasks of varying complexities remain poorly understood. In this paper, we explore the hypothesis that LLMs process concepts of varying complexities in different layers, introducing the idea of “Concept Depth” to suggest that more complex concepts are typically acquired in deeper layers. Specifically, we categorize concepts based on their level of abstraction, defining them in the order of increasing complexity within factual, emotional, and inferential tasks. We conduct extensive probing experiments using layer-wise representations across various LLM families (Gemma, LLaMA, Qwen) on various datasets spanning the three domains of tasks. Our findings reveal that models could efficiently conduct probing for simpler tasks in shallow layers, and more complex tasks typically necessitate deeper layers for accurate understanding. Additionally, we examine how external factors, such as adding noise to the input and quantizing the model weights, might affect layer-wise representations. Our findings suggest that these factors can impede the development of a conceptual understanding of LLMs until deeper layers are explored. We hope that our proposed concept and experimental insights will enhance the understanding of the mechanisms underlying LLMs. Our codes are available at https://github.com/Luckfort/CD.

Recent researches have shown that Large Language Models (LLMs) are susceptible to a security threat known as Backdoor Attack. The backdoored model will behave well in normal cases but exhibit malicious behaviours on inputs inserted with a specific backdoor trigger. Current backdoor studies on LLMs predominantly focus on single-turn instruction-tuned LLMs, while neglecting another realistic scenario where LLMs are fine-tuned on multi-turn conversational data to be chat models. Chat models are extensively adopted across various real-world scenarios, thus the security of chat models deserves increasing attention. Unfortunately, we point out that the flexible multi-turn interaction format instead increases the flexibility of trigger designs and amplifies the vulnerability of chat models to backdoor attacks. In this work, we reveal and achieve a novel backdoor attacking method on chat models by distributing multiple trigger scenarios across user inputs in different rounds, and making the backdoor be triggered only when all trigger scenarios have appeared in the historical conversations. Experimental results demonstrate that our method can achieve high attack success rates (e.g., over 90% ASR on Vicuna-7B) while successfully maintaining the normal capabilities of chat models on providing helpful responses to benign user requests. Also, the backdoor cannot be easily removed by the downstream re-alignment, highlighting the importance of continued research and attention to the security concerns of chat models. Warning: This paper may contain toxic examples.

Despite the advancements in training Large Language Models (LLMs) with alignment techniques to enhance the safety of generated content, these models remain susceptible to jailbreak, an adversarial attack method that exposes security vulnerabilities in LLMs. Notably, the Greedy Coordinate Gradient (GCG) method has demonstrated the ability to automatically generate adversarial suffixes that jailbreak state-of-the-art LLMs. However, the optimization process involved in GCG is highly time-consuming, rendering the jailbreaking pipeline inefficient. In this paper, we investigate the process of GCG and identify an issue of Indirect Effect, the key bottleneck of the GCG optimization. To this end, we propose the Model Attack Gradient Index GCG (MAGIC), that addresses the Indirect Effect by exploiting the gradient information of the suffix tokens, thereby accelerating the procedure by having less computation and fewer iterations. Our experiments on AdvBench show that MAGIC achieves up to a 1.5x speedup, while maintaining Attack Success Rates (ASR) on par or even higher than other baselines. Our MAGIC achieved an ASR of 74% on the Llama-2 and an ASR of 54% when conducting transfer attacks on GPT-3.5. Code is available at https://github.com/jiah-li/magic.

Explanation regularisation (ER) has been introduced as a way to guide text classifiers to form their predictions relying on input tokens that humans consider plausible. This is achieved by introducing an auxiliary explanation loss that measures how well the output of an input attribution technique for the model agrees with human-annotated rationales. The guidance appears to benefit performance in out-of-domain (OOD) settings, presumably due to an increased reliance on plausible tokens. However, previous work has under-explored the impact of guidance on that reliance, particularly when reliance is measured using attribution techniques different from those used to guide the model. In this work, we seek to close this gap, and also explore the relationship between reliance on plausible features and OOD performance. We find that the connection between ER and the ability of a classifier to rely on plausible features has been overstated and that a stronger reliance on plausible tokens does not seem to be the cause for OOD improvements.

The rapid pace of research publications makes it challenging for researchers to stay up to date. There is a growing need for automatically generated, concise literature reviews to help researchers quickly identify papers relevant to their interests. Prior work over the past decade has focused on summarizing individual research papers, typically in the context of citation generation, while the relationships among multiple papers have largely been overlooked. Existing approaches primarily generate standalone citation sentences without addressing the need for expository and transition sentences to explain the relationships among multiple citations. In this work, we propose a feature-based, LLM-prompting approach to generate richer citation texts and simultaneously capture the complex relationships among multiple papers. Our expert evaluation reveals a strong correlation between human preference and integrative writing styles, indicating that readers favor high-level, abstract citations with transition sentences that weave them into a coherent narrative.

Exploring effective collaboration among multiple large language models (LLMs) represents an active research direction, with multiagent debate (MAD) emerging as a popular approach. MAD involves LLMs independently generating responses and refining their own responses by incorporating feedback from other agents in a debate manner. However,empirical experiments reveal the suboptimal performance of MAD in complex reasoning scenarios. We attribute this to the potential misleading caused by peer agents with limited individual capabilities. To address this, we propose a novel sequential collaboration framework named Explain-Analyze-Generate(EAG). By decomposing complex tasks into essential subtasks and employing a pipeline approach, EAG enable agents provide constructive assistance to peers, ultimately yielding higher performance. We conduct experiments on the comprehensive complex language reasoning benchmark: BIG-Bench-Hard (BBH). Our method achieves the highest performance on 19 out of 23 tasks, with an average improvement of 8% across all tasks, and incurs lower costs compared to MAD, demonstrating its effectiveness and efficiency.

Hateful memes continuously evolve as new ones emerge by blending progressive cultural ideas, rendering existing methods that rely on extensive training obsolete or ineffective. In this work, we propose Evolver, which incorporates Large Multimodal Models (LMMs) via Chain-of-Evolution (CoE) Prompting, by integrating the evolution attribute and in-context information of memes. Specifically, Evolver simulates the evolving and expressing process of memes and reasons through LMMs in a step-by-step manner using an evolutionary pair mining module, an evolutionary information extractor, and a contextual relevance amplifier. Extensive experiments on public FHM, MAMI, and HarM datasets show that CoE prompting can be incorporated into existing LMMs to improve their performance. More encouragingly, it can serve as an interpretive tool to promote the understanding of the evolution of memes.

Large language models (LLMs) possess extensive prior knowledge and powerful in-context learning (ICL) capabilities, presenting significant opportunities for low-resource tasks. Though effective, several key issues still have not been well-addressed when focusing on zero-shot named entity recognition (NER), including the misalignment between model and human definitions of entity types, and confusion of similar types. This paper proposes an Evolving Prompts framework that guides the model to better address these issues through continuous prompt refinement. Specifically, we leverage the model to summarize the definition of each entity type and the distinctions between similar types (i.e., entity type guidelines). An iterative process is introduced to continually adjust and improve these guidelines. Additionally, since high-quality demonstrations are crucial for effective learning yet challenging to obtain in zero-shot scenarios, we design a strategy motivated by self-consistency and prototype learning to extract reliable and diverse pseudo samples from the model’s predictions. Experiments on four benchmarks demonstrate the effectiveness of our framework, showing consistent performance improvements.

Large language models (LLMs) have shown potential as general evaluators along with the evident benefits of speed and cost. While their correlation against human annotators has been widely studied, consistency as evaluators is still understudied, raising concerns about the reliability of LLM evaluators. In this paper, we conduct extensive studies on the two aspects of consistency in LLM evaluations, Self-Consistency (SC) and Inter-scale Consistency (IC), on different scoring scales and criterion granularity with open-source and proprietary models. Our comprehensive analysis demonstrates that strong proprietary models are not necessarily consistent evaluators, highlighting the importance of considering consistency in assessing the capability of LLM evaluators.

Large Language Models (LLMs) show promising learning and reasoning abilities. Compared to other NLP tasks, multilingual and multi-label emotion evaluation tasks are under-explored in LLMs. In this paper, we present EthioEmo, a multi-label emotion classification dataset for four Ethiopian languages, namely, Amharic (amh), Afan Oromo (orm), Somali (som), and Tigrinya (tir). We perform extensive experiments with an additional English multi-label emotion dataset from SemEval 2018 Task 1. Our evaluation includes encoder-only, encoder-decoder, and decoder-only language models. We compare zero and few-shot approaches of LLMs to fine-tuning smaller language models. The results show that accurate multi-label emotion classification is still insufficient even for high-resource languages such as English, and there is a large gap between the performance of high-resource and low-resource languages. The results also show varying performance levels depending on the language and model type. EthioEmo is available publicly to further improve the understanding of emotions in language models and how people convey emotions through various languages.