COLING 2025 论文列表

In the era of social media, the proliferation of fake news has created an urgent need for more effective detection methods, particularly for multimodal content. The task of identifying fake news is highly challenging, as it requires broad background knowledge and understanding across various domains. Existing detection methods primarily rely on neural networks to learn latent feature representations, resulting in black-box classifications with limited real-world understanding. To address these limitations, we propose a novel approach that leverages Multimodal Large Language Models (MLLMs) for fake news detection. Our method introduces adversarial reasoning through debates from opposing perspectives. By harnessing the powerful capabilities of MLLMs in text generation and cross-modal reasoning, we guide these models to engage in multimodal debates, generating adversarial arguments based on contradictory evidence from both sides of the issue. We then utilize these arguments to learn reasonable thinking patterns, enabling better multimodal fusion and fine-tuning. This process effectively positions our model as a debate referee for adversarial inference. Extensive experiments conducted on four fake news detection datasets demonstrate that our proposed method significantly outperforms state-of-the-art approaches.

Large language model unlearning has garnered increasing attention due to its potential to address security and privacy concerns, leading to extensive research in the field. However, existing studies have predominantly focused on instance-level unlearning, specifically targeting the removal of predefined instances containing sensitive content. This focus has left a gap in the exploration of removing an entire entity, which is critical in real-world scenarios such as copyright protection. To close this gap, we propose a novel task named Entity-level unlearning, which aims to erase entity-related knowledge from the target model completely. To investigate this task, we systematically evaluate popular unlearning algorithms, revealing that current methods struggle to achieve effective entity-level unlearning. Then, we further explore the factors that influence the performance of unlearning algorithms, identifying that the knowledge coverage of the forget set and its size play pivotal roles. Notably, our analysis also uncovers that entities introduced through fine-tuning are more vulnerable than pre-trained entities during unlearning. We hope these findings can inspire future improvements in entity-level unlearning for LLMs.

As Large Language Models (LLMs) become more advanced, the security risks they pose also increase. Ensuring that LLM behavior aligns with human values, particularly in mitigating jailbreak attacks with elusive and implicit intentions, has become a significant challenge. To address this issue, we propose a jailbreak defense method called Real Intentions Defense (RID), which involves two phases: soft extraction and hard deletion. In the soft extraction phase, LLMs are leveraged to extract unbiased, genuine intentions, while in the hard deletion phase, a greedy gradient-based algorithm is used to remove the least important parts of a sentence, based on the insight that words with smaller gradients have less impact on its meaning. We conduct extensive experiments on Vicuna and Llama2 models using eight state-of-the-art jailbreak attacks and six benchmark datasets. Our results show a significant reduction in both Attack Success Rate (ASR) and Harmful Score of jailbreak attacks, while maintaining overall model performance. Further analysis sheds light on the underlying mechanisms of our approach.

The rapid development of large language models (LLMs) has significantly improved the generation of fluent and convincing text, raising concerns about their potential misuse on social media platforms. We present a comprehensive methodology for creating nine Twitter datasets to examine the generative capabilities of four prominent LLMs: Llama 3, Mistral, Qwen2, and GPT4o. These datasets encompass four censored and five uncensored model configurations, including 7B and 8B parameter base-instruction models of the three open-source LLMs. Additionally, we perform a data quality analysis to assess the characteristics of textual outputs from human, “censored,” and “uncensored models,” employing semantic meaning, lexical richness, structural patterns, content characteristics, and detector performance metrics to identify differences and similarities. Our evaluation demonstrates that “uncensored” models significantly undermine the effectiveness of automated detection methods. This study addresses a critical gap by exploring smaller open-source models and the ramifications of “uncensoring,” providing valuable insights into how domain adaptation and content moderation strategies influence both the detectability and structural characteristics of machine-generated text.

Large Language Models (LLMs) rely on subword vocabularies to process and generate text. However, because subwords are marked as initial- or intra-word, we find that LLMs perform poorly at handling some types of affixations, which hinders their ability to generate novel (unobserved) word forms. The largest models trained on enough data can mitigate this tendency because their initial- and intra-word embeddings are aligned; in-context learning also helps when all examples are selected in a consistent way; but only morphological segmentation can achieve a near-perfect accuracy.

Entity alignment (EA) aims to match identical entities across different knowledge graphs (KGs). Graph neural network-based entity alignment methods have achieved promising results in Euclidean space. However, KGs often contain complex local and hierarchical structures, which are hard to represent in a single space. In this paper, we propose a novel method named as UniEA, which unifies dual-space embedding to preserve the intrinsic structure of KGs. Specifically, we simultaneously learn graph structure embeddings in both Euclidean and hyperbolic spaces to maximize the consistency between embeddings in the two spaces. Moreover, we employ contrastive learning to mitigate the misalignment issues caused by similar entities, where embeddings of similar neighboring entities become too close. Extensive experiments on benchmark datasets demonstrate that our method achieves state-of-the-art performance in structure-based EA. Our code is available at https://github.com/wonderCS1213/UniEA.

Grammatical Error Correction is an important research direction in NLP field. Although many models of different architectures and datasets across different languages have been developed to support the research, there is a lack of a comprehensive evaluation on these models, and different architectures make it hard for developers to implement these models on their own. To address this limitation, we present UnifiedGEC, the first open-source GEC-oriented toolkit, which consists of several core components and reusable modules. In UnifiedGEC, we integrate 5 widely-used GEC models and compare their performance on 7 datasets in different languages. Additionally, GEC-related modules such as data augmentation, prompt engineering are also deployed in it. Developers are allowed to implement new models, run and evaluate on existing benchmarks through our framework in a simple way. Code, documents and detailed results of UnifiedGEC are available at https://github.com/AnKate/UnifiedGEC.

Aspect Sentiment Quad Prediction (ASQP) aims to extract all sentiment elements in quads for a given review to explain the reason for the sentiment. Previous table-filling based methods have achieved promising results by modeling word-pair relations. However, these methods decompose the ASQP task into several subtasks without considering the association between sentiment elements. Most importantly, they fail to tackle the situation where a sentence contains multiple implicit expressions. To address these limitations, we propose a simple yet effective Unified Grid Tagging Scheme (UGTS) to extract sentiment quadruplets in one shot, with two additional special tokens from pre-trained models to represent potential implicit aspect and opinion terms. Based on this, we first introduce the adaptive graph diffusion convolution network to construct the direct connection between explicit and implicit sentiment elements from syntactic and semantic views. Next, we utilize conditional layer normalization to refine the mutual indication effect between words for matching valid aspect-opinion pairs. Finally, we employ the triaffine mechanism to integrate heterogeneous word-pair relations to capture higher-order interactions between sentiment elements. Experimental results on four benchmark datasets show the effectiveness and robustness of our model, which achieves state-of-the-art performance.

Enabling LLMs to handle lengthy context is currently a research hotspot. Most LLMs are built upon rotary position embedding (RoPE), a popular position encoding method. Therefore, a prominent path is to extrapolate the RoPE trained on comparably short texts to far longer texts. A heavy bunch of efforts have been dedicated to boosting the extrapolation via extending the formulations of the RoPE, however, few of them have attempted to showcase their inner workings comprehensively. In this paper, we are driven to offer a straightforward yet in-depth understanding of RoPE extensions from an attention perspective and on two benchmarking tasks. A broad array of experiments reveals several valuable findings: 1) Maintaining attention patterns to those at the pretrained length improves extrapolation; 2) Large attention uncertainty leads to retrieval errors; 3) Using longer continual pretraining lengths for RoPE extensions could reduce attention uncertainty and significantly enhance extrapolation.

In this paper, we investigate the output token probability information in the output embedding of language models. We find an approximate common log-linear encoding of output token probabilities within the output embedding vectors and empirically demonstrate that it is accurate and sparse. As a causality examination, we steer the encoding in output embedding to modify the output probability distribution accurately. Moreover, the sparsity we find in output probability encoding suggests that a large number of dimensions in the output embedding do not contribute to causal language modeling. Therefore, we attempt to delete the output-unrelated dimensions and find more than 30% of the dimensions can be deleted without significant movement in output distribution and sequence generation. Additionally, in the pre-training dynamics of language models, we find that the output embeddings capture the corpus token frequency information in early steps, even before an obvious convergence of parameters starts.

NLP models often face challenges with under-represented languages due to a lack of sufficient training data and language complexities. This can result in inaccurate predictions and a failure to capture the inherent uncertainties within these languages. This paper introduces a new method for modelling uncertainty in under-represented languages by employing deep Bayesian Gaussian Processes. We develop a novel framework that integrates prior knowledge and leverages kernel functions. This helps enable the quantification of uncertainty in predictions to overcome the data limitations in under-represented languages. The efficacy of our approach is validated through various experiments, and the results are benchmarked against existing methods to highlight the enhancements in prediction accuracy and measurement of uncertainty.

Online reporting platforms have enabled citizens around the world to collectively share their opinions and report in real time on events impacting their local communities. Systematically organizing (e.g., categorizing by attributes) and geotagging large amounts of crowdsourced information is crucial to ensuring that accurate and meaningful insights can be drawn from this data and used by policy makers to bring about positive change. These tasks, however, typically require extensive manual annotation efforts. In this paper we present Uchaguzi-2022, a dataset of 14k categorized and geotagged citizen reports related to the 2022 Kenyan General Election containing mentions of election-related issues such as official misconduct, vote count irregularities, and acts of violence. We use this dataset to investigate whether language models can assist in scalably categorizing and geotagging reports, thus highlighting its potential application in the AI for Social Good space.

URIEL is a knowledge base offering geographical, phylogenetic, and typological vector representations for 7970 languages. It includes distance measures between these vectors for 4005 languages, which are accessible via the lang2vec tool. Despite being frequently cited, URIEL is limited in terms of linguistic inclusion and overall usability. To tackle these challenges, we introduce URIEL+, an enhanced version of URIEL and lang2vec that addresses these limitations. In addition to expanding typological feature coverage for 2898 languages, URIEL+ improves the user experience with robust, customizable distance calculations to better suit the needs of users. These upgrades also offer competitive performance on downstream tasks and provide distances that better align with linguistic distance studies.

The two-stage retrieval paradigm has gained popularity, where a neural model serves as a re-ranker atop a non-neural first-stage retriever. We argue that this approach, involving two disparate models without interaction, represents a suboptimal choice. To address this, we propose a unified encoder-decoder architecture with a novel training regimen which enables the encoder representation to be used for retrieval and the decoder for re-ranking within a single unified model, facilitating end-to-end retrieval. We incorporate XTR-style retrieval on top of the trained MonoT5 reranker to specifically concentrate on addressing practical constraints to create a lightweight model. Results on the BIER benchmark demonstrate the effectiveness of our unified architecture, featuring a highly optimized index and parameters. It outperforms ColBERT, XTR, and even serves as a superior re-ranker compared to the Mono-T5 reranker. The performance gains of our proposed system in reranking become increasingly evident as model capacity grows, particularly when compared to rerankers operating over traditional first-stage retrievers like BM25. This is encouraging, as it suggests that we can integrate more advanced retrievers to further enhance final reranking performance. In contrast, BM25’s static nature limits its potential for such improvements.

In the field of natural language generation (NLG), controlling text generation (CTG) is critical, particularly in query auto-completion (QAC) where the need for personalization and diversity is paramount. However, it is essentially challenging to adapt to various control objectives and constraints, which results in existing CTG approaches meeting with mixed success. This paper presents UCTG, a unified controllable text generation framework, which introduces a novel prompt learning method for CTG. Specifically, this framework seamlessly integrates a control module, a prompt module, and a generation module. The control module leverages a fine-tuned model to distill user preference features and behavioral patterns from historical data, incorporating human feedback into the model’s loss functions. These features are then transformed by the prompt module into vectors that guide the generation module. As such, the text generation can be flexibly controlled without modifying the task settings. By employing this unified approach, UCTG significantly improves query accuracy and coherence in tasks with different objectives and constraints, which is validated by extensive experiments on the Meituan and AOL real-world datasets. UCTG not only improves text generation control in QAC but also sets a new framework for flexible NLG applications.

Previous work on Incomplete Utterance Rewriting (IUR) has primarily focused on generating rewritten utterances based solely on dialogue context, ignoring the widespread phenomenon of coreference and ellipsis in dialogues. To address this issue, we propose a novel framework called TEO (Two-stage approach on Editing Operation) for IUR, in which the first stage generates editing operations and the second stage rewrites incomplete utterances utilizing the generated editing operations and the dialogue context. Furthermore, an adversarial perturbation strategy is proposed to mitigate cascading errors and exposure bias caused by the inconsistency between training and inference in the second stage. Experimental results on three IUR datasets show that our TEO outperforms the SOTA models significantly.

Document-level event factuality identification (DEFI) assesses the veracity degree to which an event mentioned in a document has happened, which is crucial for many natural language processing tasks. Previous work assesses event factuality by solely relying on the semantic information within a single document, which fails to identify hard cases where the document itself is hallucinative or counterfactual. There is also a pressing need for more suitable data of this kind. To tackle these issues, we construct Factualusion, a novel corpus with hallucination features that can be used not only for DEFI but can also be applied for hallucination evaluation for large language models. We further propose Trucidator, a graph-based framework that constructs intra-document and cross-document graphs and employs a multi-task learning paradigm to acquire more robust node embeddings, leveraging cross-document inference for more accurate identification. Experiments show that our proposed framework outperformed several baselines, demonstrating the effectiveness of our method.

Current video-guided machine translation (VMT) approaches primarily use coarse-grained visual information, resulting in information redundancy, high computational overhead, and neglect of audio content. Our research demonstrates the significance of fine-grained visual and audio information in VMT from both data and methodological perspectives. From the data perspective, we have developed a large-scale dataset TriFine, the first vision-audio-subtitle tri-modal VMT dataset with annotated multimodal fine-grained tags. Each entry in this dataset not only includes the triples found in traditional VMT datasets but also encompasses seven fine-grained annotation tags derived from visual and audio modalities. From the methodological perspective, we propose a Fine-grained Information-enhanced Approach for Translation (FIAT). Experimental results have shown that, in comparison to traditional coarse-grained methods and text-only models, our fine-grained approach achieves superior performance with lower computational overhead. These findings underscore the pivotal role of fine-grained annotated information in advancing the field of VMT.

A domain-specific extension of C language known as extended Berkeley Packet Filter (eBPF) has gained widespread acceptance for various tasks, including observability, security, and network acceleration in the cloud community. Due to its recency and complexity, there is an overwhelming need for natural language summaries of existing eBPF codes (particularly open-source code) for practitioners and developers, which will go a long way in easing the understanding and development of new code. However, being a niche Domain-Specific Language (DSL), there is a scarcity of available training data. In this paper, we investigate the effectiveness of LLMs for summarizing low-resource DSLs, in the context of eBPF codes. Specifically, we propose a clustering-based technique to retrieve in-context examples that are semantically closer to the test example and propose a very simple yet powerful prompt design that yields superior-quality code summary generation. Experimental results show that our proposed retrieval approach for prompt generation improves the eBPF code summarization accuracy up to 12.9 BLEU points over other prompting techniques. The codes are available at https://github.com/babangain/ebpf_summ.