COLING 2025 Papers

This paper focuses on sarcasm detection, which aims to identify whether given statements convey criticism, mockery, or other negative sentiment opposite to the literal meaning. To detect sarcasm, humans often require a comprehensive understanding of the semantics in the statement and even resort to external commonsense to infer the fine-grained incongruity. However, existing methods lack commonsense inferential ability when they face complex real-world scenarios, leading to unsatisfactory performance. To address this problem, we propose a novel framework for sarcasm detection, which conducts incongruity reasoning based on commonsense augmentation, called EICR. Concretely, we first employ retrieval-augmented large language models to supplement the missing but indispensable commonsense background knowledge. To capture complex contextual associations, we construct a dependency graph and obtain the optimized topology via graph refinement. We further introduce an adaptive reasoning skeleton that integrates prior rules to extract sentiment-inconsistent subgraphs explicitly. To eliminate the possible spurious relations between words and labels, we employ adversarial contrastive learning to enhance the robustness of the detector. Experiments conducted on five datasets demonstrate the effectiveness of EICR.

Mental manipulation severely undermines mental wellness by covertly and negatively distorting decision-making. While there is an increasing interest in mental health care within the natural language processing community, progress in tackling manipulation remains limited due to the complexity of detecting subtle, covert tactics in conversations. In this paper, we propose Intent-Aware Prompting (IAP), a novel approach for detecting mental manipulations using large language models (LLMs), providing a deeper understanding of manipulative tactics by capturing the underlying intents of participants. Experimental results on the MentalManip dataset demonstrate superior effectiveness of IAP against other advanced prompting strategies. Notably, our approach substantially reduces false negatives, helping detect more instances of mental manipulation with minimal misjudgment of positive cases. The code of this paper is available at https://github.com/Anton-Jiayuan-MA/Manip-IAP.

In the industry, numerous tasks are deployed online. Traditional approaches often tackle each task separately by its own network, which leads to excessive costs for developing and scaling models, especially in the context of large language models. Although multi-task methods can save costs through parameter sharing, they often struggle to outperform single-task methods in real-world applications. To tackle these challenges, we present a three-stage multi-task learning framework for large language models. It involves task filtering, followed by fine-tuning on high-resource tasks, and finally fine-tuning on all tasks. We conducted comprehensive experiments in single-task and multi-task settings. Our approach, exemplified on different benchmarks, demonstrates that it is able to achieve performance comparable to the single-task method while reducing up to 90.9% of its overhead.

When building a predictive model, it is often difficult to ensure that application-specific requirements are encoded by the model that will eventually be deployed. Consider researchers working on hate speech detection. They will have an idea of what is considered hate speech, but building a model that reflects their view accurately requires preserving those ideals throughout the workflow of data set construction and model training. Complications such as sampling bias, annotation bias, and model misspecification almost always arise, possibly resulting in a gap between the application specification and the model’s actual behavior upon deployment. To address this issue for hate speech detection, we propose DefVerify: a 3-step procedure that (i) encodes a user-specified definition of hate speech, (ii) quantifies to what extent the model reflects the intended definition, and (iii) tries to identify the point of failure in the workflow. We use DefVerify to find gaps between definition and model behavior when applied to six popular hate speech benchmark datasets.

Theory of Mind (ToM) is the ability to understand and reflect on the mental states of others. Although this capability is crucial for human interaction, testing on Large Language Models (LLMs) reveals that they possess only a rudimentary understanding of it. Although the most capable closed-source LLMs have come close to human performance on some ToM tasks, they still perform poorly on complex variations of the task that involve more structured reasoning. In this work, we utilize the concept of “pretend-play”, or “Simulation Theory” from cognitive psychology to propose “Decompose-ToM”: an LLM-based inference algorithm that improves model performance on complex ToM tasks. We recursively simulate user perspectives and decompose the ToM task into a simpler set of tasks: subject identification, question-reframing, world model updation, and knowledge availability. We test the algorithm on higher-order ToM tasks and a task testing for ToM capabilities in a conversational setting, demonstrating that our approach shows significant improvement across models compared to baseline methods while requiring minimal prompt tuning across tasks and no additional model training. Our code is publicly available.

The impact of social media on critical issues such as echo chambers, needs to be addressed, as these phenomena can have disruptive consequences for our society. Traditional research often oversimplifies emotional tendencies and opinion evolution into numbers and formulas, neglecting that news and communication are conveyed through text, which limits these approaches. Hence, in this work, we propose an LLM-based simulation for the social opinion network to evaluate and counter polarization phenomena. We first construct three typical network structures to simulate different characteristics of social interactions. Then, agents interact based on recommendation algorithms and update their strategies through reasoning and analysis. By comparing these interactions with the classic Bounded Confidence Model (BCM), the Friedkin-Johnsen (FJ) model, and using echo chamber-related indices, we demonstrate the effectiveness of our framework in simulating opinion dynamics and reproducing phenomena such as opinion polarization and echo chambers. We propose two mitigation methods—active and passive nudges—that can help reduce echo chambers, specifically within language-based simulations. We hope our work will offer valuable insights and guidance for social polarization mitigation.

Decoding strategies for generative large language models (LLMs) are a critical but often underexplored aspect of text generation tasks. Guided by specific hyperparameters, these strategies aim to transform the raw probability distributions produced by language models into coherent, fluent text. In this study, we undertake a large-scale empirical assessment of a range of decoding methods, open-source LLMs, textual domains, and evaluation protocols to determine how hyperparameter choices shape the outputs. Our experiments include both factual (e.g., news) and creative (e.g., fiction) domains, and incorporate a broad suite of automatic evaluation metrics alongside human judgments. Through extensive sensitivity analyses, we distill practical recommendations for selecting and tuning hyperparameters, noting that optimal configurations vary across models and tasks. By synthesizing these insights, this study provides actionable guidance for refining decoding strategies, enabling researchers and practitioners to achieve higher-quality, more reliable, and context-appropriate text generation outcomes.

While there has been considerable amount of research on bias mitigation algorithms, two properties: multi-community perspective and fairness to *all* communities have not been given sufficient attention. Focusing on these, we propose an obfuscation based data augmentation debiasing approach. In it we add to the training data *obfuscated* versions of *all* false positive instances irrespective of source community. We test our approach by debiasing toxicity classifiers built using 5 neural models (multi layer perceptron model and masked language models) and 3 datasets in a 4 communities setting. We also explore 4 different obfuscators for debiasing. Results demonstrate the merits of our approach: bias is reduced for almost all of our runs without sacrificing false positive rates or F1 scores for minority or majority communities. In contrast, the 4 state of the art baselines typically make performance sacrifices (often large) while reducing bias. Crucially, we demonstrate that it is possible to debias while maintaining standards for both minority and majority communities.

Writing arguments is a challenging task for both humans and machines. It entails incorporating high-level beliefs from various perspectives on the topic, along with deliberate reasoning and planning to construct a coherent narrative. Current language models often generate outputs autoregressively, lacking explicit integration of these underlying controls, resulting in limited output diversity and coherence. In this work, we propose a persona-based multi-agent framework for argument writing. Inspired by the human debate, we first assign each agent a persona representing its high-level beliefs from a unique perspective, and then design an agent interaction process so that the agents can collaboratively debate and discuss the idea to form an overall plan for argument writing. Such debate process enables fluid and nonlinear development of ideas. We evaluate our framework on argumentative essay writing. The results show that our framework generates more diverse and persuasive arguments by both automatic and human evaluations.

While in-context Learning (ICL) has proven to be an effective technique to improve the performance of Large Language Models (LLMs) in a variety of complex tasks, notably in translating natural language questions into Structured Query Language (NL2SQL), the question of how to select the most beneficial demonstration examples remains an open research problem. While prior works often adapted off-the-shelf encoders to retrieve examples dynamically, an inherent discrepancy exists in the representational capacities between the external retrievers and the LLMs. Further, optimizing the selection of examples is a non-trivial task, since there are no straightforward methods to assess the relative benefits of examples without performing pairwise inference. To address these shortcomings, we propose Detriever, a novel demonstration retrieval framework that learns a weighted combination of LLM hidden states, where rich semantic information is encoded. To train the model, we propose a proxy score that estimates the relative benefits of examples based on the similarities between output queries. Experiments on two popular NL2SQL benchmarks demonstrate that our method significantly outperforms the state-of-the-art baselines for the NL2SQL tasks.

In recent years, the use of large language models (LLMs) for text classification has attracted widespread attention. Despite this, the classification accuracy of LLMs has not yet universally surpassed that of smaller models. LLMs can enhance their performance in text classification through fine-tuning. However, existing data quality research based on LLMs is challenging to apply directly to solve text classification problems. To further improve the performance of LLMs in classification tasks, this paper proposes a data quality enhancement (DQE) method for text classification based on LLMs. This method starts by using a greedy algorithm to select data, dividing the dataset into sampled and unsampled subsets, and then performing fine-tuning of the LLMs using the sampled data. Subsequently, this model is used to predict the outcomes for the unsampled data, categorizing incorrectly predicted data into uncovered, difficult, and noisy data. Experimental results demonstrate that our method effectively enhances the performance of LLMs in text classification tasks and significantly improves training efficiency, saving nearly half of the training time. Our method has achieved state-of-the-art performance in several open-source classification tasks.

Cross-domain constituency parsing remains a challenging task due to the lack of high-quality out-of-domain data. In this paper, we propose a data augmentation method via lightweight large language model (LLM) generation and tree hybridization. We utilize LLM to generate phrase structures (subtrees) for the target domain by incorporating grammar rules and lexical head information into the prompt. To better leverage LLM-generated target-domain subtrees, we hybridize them with existing source-domain subtrees to efficiently produce a large number of structurally diverse instances. Experimental results demonstrate that our method achieves significant improvements on five target domains with a lightweight LLM generation cost.

Translation of low-resource languages in industrial domains is essential for improving market productivity and ensuring foreign workers have better access to information. However, existing translators struggle with domain-specific terms, and there is a lack of expert annotators for dataset creation. In this work, we propose DaCoM, a methodology for collecting low-resource language pairs from industrial domains to address these challenges. DaCoM is a hybrid translation framework enabling effective data collection. The framework consists of a large language model and neural machine translation. Evaluation verifies existing models perform inadequately on DaCoM-created datasets, with up to 53.7 BLEURT points difference depending on domain inclusion. DaCoM is expected to address the lack of datasets for domain-specific low-resource languages by being easily pluggable into future state-of-the-art models and maintaining an industrial domain-agnostic approach.

The susceptibility of deep neural networks to adversarial attacks is a well-established concern. To address this problem, robustness certification is proposed, which, unfortunately, suffers from precision or scalability issues. In this paper, we present DROWN (Dual CROWN), a novel method for certifying the robustness of DNNs. The advantage of DROWN is that it tightens classic LiRPA-based methods yet maintains similar scalability, which comes from refining pre-activation bounds of ReLU relaxations using two pairs of linear bounds derived from different relaxations of ReLU units in previous layers. The extensive evaluations show that DROWN achieves up to 83.39% higher certified robust accuracy than the baseline on CNNs and up to 4.68 times larger certified radii than the baseline on Transformers. Meanwhile, the running time of DROWN is about twice that of the baseline.

Training models with differential privacy has received a lot of attentions since differential privacy provides theoretical guarantee of privacy preservation. For a task in a specific domain, since a large-scale pre-trained model in the same domain contains general knowledge of the task, using such a model requires less effort in designing and training the model. However, differentially privately fine-tuning such models having a large number of trainable parameters results in large degradation of utility. Thus, we propose methods that effectively fine-tune the large-scale pre-trained models with freezing unimportant parameters for downstream tasks while satisfying differential privacy. To select the parameters to be fine-tuned, we propose several efficient methods based on the gradients of model parameters. We show the effectiveness of the proposed method by performing experiments with real datasets.

Autonomous agents powered by large language models (LLMs) hold significant potential across various domains. The Reflection framework is designed to help agents learn from past mistakes in complex tasks. While previous research has shown that reflection can enhance performance, our investigation reveals a key limitation: meaningful self-reflection primarily occurs at the beginning of iterations, with subsequent attempts failing to produce further improvements. We term this phenomenon “Early Stop Reflection,” where the reflection process halts prematurely, limiting the agent’s ability to engage in continuous learning. To address this, we propose the DORA method (Dynamic and Optimized Reflection Advice), which generates task-adaptive and diverse reflection advice. DORA introduces an external open-source small language model (SLM) that dynamically generates prompts for the reflection LLM. The SLM uses feedback from the agent and optimizes the prompt generation process through a non-gradient Bayesian Optimization (BO) algorithm, ensuring the reflection process evolves and adapts over time. Our experiments in the MiniWoB++ and Alfworld environments confirm that DORA effectively mitigates the “Early Stop Reflection” issue, enabling agents to maintain iterative improvements and boost performance in long-term, complex tasks. Code are available at https://anonymous.4open.science/r/DORA-44FB/.

Recent advancements in event argument extraction (EAE) involve incorporating useful auxiliary information into models during training and inference, such as retrieved instances and event templates. These methods face two challenges: (1) the retrieval results may be irrelevant and (2) templates are developed independently for each event without considering their possible relationship. In this work, we propose DEGAP to address these challenges through a simple yet effective components: dual prefixes, i.e. learnable prompt vectors, where the instance-oriented prefix and template-oriented prefix are trained to learn information from different event instances and templates. Additionally, we propose an event-guided adaptive gating mechanism, which can adaptively leverage possible connections between different events and thus capture relevant information from the prefix. Finally, these event-guided prefixes provide relevant information as cues to EAE model without retrieval. Extensive experiments demonstrate that our method achieves new state-of-the-art performance on four datasets (ACE05, RAMS, WIKIEVENTS, and MLEE). Further analysis shows the impact of different components.

Entity Alignment (EA) is a critical task in Knowledge Graph (KG) integration, aimed at identifying and matching equivalent entities that represent the same real-world objects. While EA methods based on knowledge representation learning have shown strong performance on synthetic benchmark datasets such as DBP15K, their effectiveness significantly decline in real-world scenarios which often involve data that is highly heterogeneous, incomplete, and domain-specific, as seen in datasets like DOREMUS and AGROLD. Addressing this challenge, we propose DAEA, a novel EA approach with Domain Adaptation that leverages the data characteristics of synthetic benchmarks for improved performance in real-world datasets. DAEA introduces a multi-source KGs selection mechanism and a specialized domain adaptive entity alignment loss function to bridge the gap between real-world data and optimal benchmark data, mitigating the challenges posed by aligning entities across highly heterogeneous KGs. Experimental results demonstrate that DAEA outperforms state-of-the-art models on real-world datasets, achieving a 29.94% improvement in Hits@1 on DOREMUS and a 5.64% improvement on AGROLD. Code is available at https://github.com/yangxiaoxiaoly/DAEA.

Open Information Extraction (OpenIE) aims to extract structured information in the form of triples from unstructured text, serving as a foundation for various downstream NLP tasks. Despite the success of neural OpenIE models, their dependence on large-scale annotated datasets poses a challenge, particularly in low-resource settings. In this paper, we introduce a novel approach to address the low-resource OpenIE task through two key innovations: (1) we improve the quality of training data by curating small-scale, high-quality datasets annotated by a large language model (GPT-3.5), leveraging both OpenIE principles and few-shot examples to form LSOIE-g principles and LSOIE-g examples; (2) we propose CycleOIE, a training framework that maximizes data efficiency through a cycle-consistency mechanism, enabling the model to learn effectively from minimal data. Experimental results show that CycleOIE, when trained on only 2k+ instances, achieves comparable results to models trained on over 90k instances. Our contributions are further validated through extensive experiments, demonstrating the superior performance of CycleOIE and our curated LSOIE-g datasets in low-resource OpenIE as well as revealing the internal mechanisms of CycleOIE.