COLING 2025 论文列表

Multimodal hate detection aims to identify hate content across multiple modalities for promoting a harmonious online environment. Despite promising progress, three critical challenges, the absence of implicit hateful cues, the cross-modal-induced hate, and the diversity of hate target groups, inherent in the multimodal hate detection task, have been overlooked. To address these challenges, we propose a hypergraph-based prompting fusion model. Our model first uses tailored prompts to infer implicit hateful cues. It then introduces hyperedges to capture cross-modal-induced hate and applies a diversity-oriented hyperedge expansion strategy to account for different hate target groups. Finally, hypergraph convolution fuses diverse hateful cues, enhancing the exploration of cross-modal hate and targeting specific groups. Experimental results on two benchmark datasets show that our model achieves state-of-the-art performance in multimodal hate detection.

This paper introduces the human-like embodied AI interviewer which integrates android robots equipped with advanced conversational capabilities, including attentive listening, conversational repairs, and user fluency adaptation. Moreover, it can analyze and present results post-interview. We conducted a real-world case study at SIGDIAL 2024 with 42 participants, of whom 69% reported positive experiences. This study demonstrated the system’s effectiveness in conducting interviews just like a human and marked the first employment of such a system at an international conference. The demonstration video is available at https://youtu.be/jCuw9g99KuE.

Human Interest (HI) framing is a narrative strategy that injects news stories with a relatable, emotional angle and a human face to engage the audience. In this study we investigate the use of HI framing across different English-speaking cultures in news articles about climate change. Despite its demonstrated impact on the public’s behaviour and perception of an issue, HI framing has been under-explored in NLP to date. We perform a systematic analysis of HI stories to understand its role in climate change reporting in English-speaking countries from four continents. Our findings reveal key differences in how climate change is portrayed across countries, encompassing aspects such as narrative roles, article polarity, pronoun prevalence, and topics. We also demonstrate that these linguistic aspects boost the performance of fine-tuned pre-trained language models on HI story classification.

Within numerical reasoning, understanding numbers themselves is still a challenge for existing language models. Simple generalisations, such as solving 100+200 instead of 1+2, can substantially affect model performance (Sivakumar and Moosavi, 2023). Among various techniques, character-level embeddings of numbers have emerged as a promising approach to improve number representation. However, this method has limitations as it leaves the task of aggregating digit representations to the model, which lacks direct supervision for this process. In this paper, we explore the use of mathematical priors to compute aggregated digit embeddings and explicitly incorporate these aggregates into transformer models. This can be achieved either by adding a special token to the input embeddings or by introducing an additional loss function to enhance correct predictions. We evaluate the effectiveness of incorporating this explicit aggregation, analysing its strengths and shortcomings, and discuss future directions to better benefit from this approach. Our methods, while simple, are compatible with any pretrained model, easy to implement, and have been made publicly available.

Long sequences occur in abundance within real-world scenarios, hence properly modelling them opens numerous down-stream use-cases. Deep neural networks, however, have often struggled with these for a variety of reasons. Recent advances, both in system engineering as well as model design, have enabled the scaling up of model that are purported to support extended context length. In particular, the state-space and linear recurrent neural network families of models hypothetically can entend to infinite sequence length. However, is this too good to be true? We conduct an evaluation to show that while such claims may be sound theoretically, there remain large practical gaps that are empirically observed. In particular, recurrent models still suffer in the same settings as long-context LLMs with attention. We further show that different inductive biases have inconsistent extrapolation capabilities, highlighting the need to further study such paradigms and investigate why long-context models seemingly fail to behave as one might expect.

Motivational Interviewing (MI) is a counseling technique that promotes behavioral change through reflective responses to mirror or refine client statements. While advanced Large Language Models (LLMs) can generate engaging dialogues, challenges remain for applying them in a sensitive context such as MI. This work assesses the potential of LLMs to generate MI reflections via three LLMs: GPT-4, Llama-2, and BLOOM, and explores the effect of dialogue context size and integration of MI strategies for reflection generation by LLMs. We conduct evaluations using both automatic metrics and human judges on four criteria: appropriateness, relevance, engagement, and naturalness, to assess whether these LLMs can accurately generate the nuanced therapeutic communication required in MI. While we demonstrate LLMs’ potential in generating MI reflections comparable to human therapists, content analysis shows that significant challenges remain. By identifying the strengths and limitations of LLMs in generating empathetic and contextually appropriate reflections in MI, this work contributes to the ongoing dialogue in enhancing LLM’s role in therapeutic counseling.

Recent studies have shown that post-aligning multilingual pretrained language models (mPLMs) using alignment objectives on both original and transliterated data can improve crosslingual alignment. This improvement further leads to better crosslingual transfer performance. However, it remains unclear how and why a better crosslingual alignment is achieved, as this technique only involves transliterations, and does not use any parallel data. This paper attempts to explicitly evaluate the crosslingual alignment and identify the key elements in transliteration-based approaches that contribute to better performance. For this, we train multiple models under varying setups for two pairs of related languages: (1) Polish and Ukrainian and (2) Hindi and Urdu. To assess alignment, we define four types of similarities based on sentence representations. Our experimental results show that adding transliterations alone improves the overall similarities, even for random sentence pairs. With the help of auxiliary transliteration-based alignment objectives, especially the contrastive objective, the model learns to distinguish matched from random pairs, leading to better crosslingual alignment. However, we also show that better alignment does not always yield better downstream performance, suggesting that further research is needed to clarify the connection between alignment and performance. The code implementation is based on https://github.com/cisnlp/Transliteration-PPA.

Instruction tuning a large language model with multiple languages can prepare it for multilingual downstream tasks. Nonetheless, it is yet to be determined whether having a handful of languages is sufficient, or whether the benefits increase with the inclusion of more. By fine-tuning large multilingual models on 1 to 52 languages, we present a case study on BLOOM to understand three pertinent factors affecting performance: the number of languages, language exposure, and similarity between training and test languages. Overall we found that 1) expanding language coverage in multilingual instruction tuning proves to be beneficial; 2) accuracy often significantly boots if the test language appears in the instruction mixture; 3) languages’ genetic features correlate with cross-lingual transfer more than merely the number of language but different languages benefit to various degrees.

Chain-of-thought emerges as a promising technique for eliciting reasoning capabilities from Large Language Models (LLMs). However, it does not always improve task performance or accurately represent reasoning processes, leaving unresolved questions about its usage. In this paper, we diagnose the underlying mechanism by comparing the reasoning process of LLMs with humans, using causal analysis to understand the relationships between the problem instruction, reasoning, and the answer in LLMs. Our empirical study reveals that LLMs often deviate from the ideal causal chain, resulting in spurious correlations and potential consistency errors (inconsistent reasoning and answers). We also examine various factors influencing the causal structure, finding that in-context learning with examples strengthens it, while post-training techniques like supervised fine-tuning and reinforcement learning on human feedback weaken it. To our surprise, the causal structure cannot be strengthened by enlarging the model size only, urging research on new techniques. We hope that this preliminary study will shed light on understanding and improving the reasoning process in LLM.

Retrieval-augmented Large Language Models (RaLLMs) are reshaping knowledge acquisition, offering long-form, knowledge-grounded answers through advanced reasoning and generation capabilities. Despite the emergence of impactful systems like WebGPT and New Bing, the reliability of RaLLMs, especially in complex situations, is under scrutiny. Our study tackles this concern by evaluating RaLLMs’ question-answering performance using a novel benchmark focusing on Correctness and Groundedness. Correctness measures the logical soundness of the responses, and Groundedness checks for support by relevant references. We introduce an automated model-based evaluation pipeline for multi-hop question-answering tasks, revealing RaLLMs’ proneness to generating inaccuracies when dealing with flawed or partial knowledge. To improve accuracy, we introduce two reasoning strategies, Self-Reflection’ and Self-Completion,’ enabling RaLLMs to identify and fill knowledge gaps, significantly improving answer quality without extensive model retraining.

The quality of rationales is essential in the reasoning capabilities of language models. Rationales not only enhance reasoning performance in complex natural language tasks but also justify model decisions. However, obtaining impeccable rationales is often impossible. Our study aims to investigate how ambiguous rationales play in model performances of natural language reasoning. We first assess the ambiguity of rationales through the lens of entropy and uncertainty in model prior beliefs, exploring its impact on task performance. We then propose a simple way to guide models to choose between two different reasoning paths depending on the ambiguity of the rationales. Our empirical results demonstrate that this approach leads to robust performance, particularly in adversarial scenarios where rationale quality is inconsistent.

Ensembling various LLMs to unlock their complementary potential and leverage their individual strengths is highly valuable. Previous studies typically focus on two main paradigms: sample-level and token-level ensembles. Sample-level ensemble methods either select or blend fully generated outputs, which hinders dynamic correction and enhancement of outputs during the generation process. On the other hand, token-level ensemble methods enable real-time correction through fine-grained ensemble at each generation step. However, the information carried by an individual token is quite limited, leading to suboptimal decisions at each step. To address these issues, we propose SweetSpan, a span-level ensemble method that effectively balances the need for real-time adjustments and the information required for accurate ensemble decisions. Our approach involves two key steps: First, we have each candidate model independently generate candidate spans based on the shared prefix. Second, we calculate perplexity scores to facilitate mutual evaluation among the candidate models and achieve robust span selection by filtering out unfaithful scores. To comprehensively evaluate ensemble methods, we propose a new challenging setting (ensemble models with significant performance gaps) in addition to the standard setting (ensemble the best-performing models) to assess the performance of model ensembles in more realistic scenarios. Experimental results in both standard and challenging settings across various language generation tasks demonstrate the effectiveness, robustness, and versatility of our approach compared with previous ensemble methods.

With the impressive performance in various downstream tasks, large language models (LLMs) have been widely integrated into production pipelines, such as recruitment and recommendation systems. A known issue of models trained on natural language data is the presence of human biases, which can impact the fairness of the system. This paper investigates LLMs’ behavior with respect to gender stereotypes in the context of occupation decision making. Our framework is designed to investigate and quantify the presence of gender stereotypes in LLMs’ behavior via multi-round question answering. Inspired by prior work, we constructed a dataset using a standard occupation classification knowledge base released by authoritative agencies. We tested it on three families of LMs (RoBERTa, GPT, and Llama) and found that all models exhibit gender stereotypes analogous to human biases, but with different preferences. The distinct preferences of GPT-3.5-turbo and Llama2-70b-chat, along with additional analysis indicating GPT-4o-mini favors female subjects, may imply that the current alignment methods are insufficient for debiasing and could introduce new biases contradicting the traditional gender stereotypes. Our contribution includes a 73,500 prompts dataset constructed with a taxonomy of real-world occupations and a multi-step verification framework to evaluate model’s behavior regarding gender stereotype.

Automated Grammatical Error Correction (GEC) has been extensively researched in Natural Language Processing (NLP), primarily focusing on English and other resource-rich languages. This paper shifts the focus to GEC for a scarcely explored low-resource language, specifically Hindi, which presents unique challenges due to its intricate morphology and complex syntax. To address data resource limitations, this work explores various GEC data generation techniques. Our research introduces a carefully extracted and filtered, high-quality dataset, HiWikiEdits, which includes human-edited 8,137 instances sourced from Wikipedia, encompassing 17 diverse grammatical error types, with annotations performed using the ERRANT toolkit. Furthermore, we investigate Round Trip Translation (RTT) using diverse languages for synthetic Hindi GEC data generation, revealing that leveraging high-resource linguistically distant language for error generation outperforms mid-resource linguistically closer languages. Specifically, using English as a pivot language resulted in a 6.25% improvement in GLEU score compared to using Assamese or Marathi. Finally, we also investigate the neural model-based synthetic error-generation technique and show that it achieves comparable performance to other synthetic data generation methods, even in low-resource settings.

Hallucination in large language models (LLMs) remains a significant challenge for their safe deployment, particularly due to its potential to spread misinformation. Most existing solutions address this challenge by focusing on aligning the models with credible sources or by improving how models communicate their confidence (or lack thereof) in their outputs. While these measures may be effective in most contexts, they may fall short in scenarios requiring more nuanced approaches, especially in situations where access to accurate data is limited or determining credible sources is challenging. In this study, we take North Korea - a country characterised by an extreme lack of reliable sources and the prevalence of sensationalist falsehoods - as a case study. We explore and evaluate how some of the best-performing multilingual LLMs and specific language-based models generate information about North Korea in three languages spoken in countries with significant geo-political interests: English (United States, United Kingdom), Korean (South Korea), and Mandarin Chinese (China). Our findings reveal significant differences, suggesting that the choice of model and language can lead to vastly different understandings of North Korea, which has important implications given the global security challenges the country poses.

Temporal knowledge graph (TKG) reasoning has become a hot topic due to its great value in many practical tasks. The key to TKG reasoning is modeling the structural information and evolutional patterns of the TKGs. While great efforts have been devoted to TKG reasoning, the structural and evolutional characteristics of real-world networks have not been considered. In the aspect of structure, real-world networks usually exhibit clear community structure and scale-free (long-tailed distribution) properties. In the aspect of evolution, the impact of an event decays with the time elapsing. In this paper, we propose a novel TKG reasoning model called Hawkes process-based Evolutional Representation Learning Network (HERLN), which learns structural information and evolutional patterns of a TKG simultaneously, considering the characteristics of real-world networks: community structure, scale-free and temporal decaying. First, we find communities in the input TKG to make the encoding get more similar intra-community embeddings. Second, we design a Hawkes process-based relational graph convolutional network to cope with the event impact-decaying phenomenon. Third, we design a conditional decoding method to alleviate biases towards frequent entities caused by long-tailed distribution. Experimental results show that HERLN achieves significant improvements over the state-of-the-art models.

Nowadays, hate speech technologies are surely relevant in Brazil. Nevertheless, the inability of these technologies to provide reasons (rationales) for their decisions is the limiting factor to their adoption since they comprise bias, which may perpetuate social inequalities when propagated at scale. This scenario highlights the urgency of proposing explainable technologies to address hate speech. However, explainable models heavily depend on data availability with human-annotated rationales, which are scarce, especially for low-resource languages. To fill this relevant gap, we introduce HateBRXplain, the first benchmark dataset for hate speech detection in Portuguese, with text span annotations capturing rationales. We evaluated our corpus using mBERT, BERTimbau, DistilBERTimbau, and PTT5 models, which outperformed the current baselines. We further assessed these models’ explainability using model-agnostic explanation methods (LIME and SHAP). Results demonstrate plausible post-hoc explanations when compared to human annotations. However, the best-performing hate speech detection models failed to provide faithful rationales.

Instruction-guided image editing consists in taking an image and an instruction and delivering that image altered according to that instruction. State-of-the-art approaches to this task suffer from the typical scaling up and domain adaptation hindrances related to supervision as they eventually resort to some kind of task-specific labelling, masking or training. We propose a novel approach that does without any such task-specific supervision and offers thus a better potential for improvement. Its assessment demonstrates that it is highly effective, achieving very competitive performance.

In light of the inherent entailment relations between images and text, embedding point vectors in hyperbolic space has been employed to leverage its hierarchical modeling advantages for visual semantic representation learning. However, point vector embeddings struggle to address semantic uncertainty, where an image may have multiple interpretations, and text may correspond to different images—a challenge especially prevalent in the medical domain. Therefor, we propose HYDEN, a novel hyperbolic density embedding based image-text representation learning approach tailored for specific medical domain data. This method integrates text-aware local features with global features from images, mapping image-text features to density features in hyperbolic space via using hyperbolic pseudo-Gaussian distributions. An encapsulation loss function is employed to model the partial order relations between image-text density distributions. Experimental results demonstrate the interpretability of our approach and its superior performance compared to the baseline methods across various zero-shot tasks and fine-tuning task on different datasets.