This paper introduces the approach we adopted for the SemEval-2025 “Food Hazard Detection” task, which aims to predict coarse-grained categories (such as “product category” and “hazard category”) and fine-grained vectors (such as specific products like “ice cream” or hazards like “salmonella”) from noisy, long-tailed text data.To address the issues of dirty data, as well as the severe long-tail distribution of text labels and length in the data, we proposed a pipeline system. This system combines data cleaning, LLM-based enhancement, label resampling, and ensemble learning to tackle data sparsity and label imbalance problems.The two subtasks have strong semantic relatedness. By integrating them into a unified multiturn dialogue framework, we fine-tuned five models using a bagging approach. Ultimately, we achieved good results in both subtasks, ranking 5th (with an F1 score of 80.17% for ST1 and 52.66% for ST2).

This paper presents our approaches for three subtasks in SemEval-2025 Task 10, which focus on entity framing, narrative classification, and narrative extraction in new analysis respectively. We propose a two-stage news analytical framework for both Subtask A and B. In Subtask A (Entity Framing), we design an entity-oriented data processing pipeline to address the issue of redundant information in a news article, and explore effective use of multilingual datasets through sufficient experiments. The system achieves the first place in Bulgarian and the second place in English and Portuguese. In Subtask B (Narrative Classification), a similar narrative-oriented data processing pipeline is adopted to obtain condensed news chunks for each narrative. We conduct in-depth discussion regarding approaches to enhancing both data quality and volume, and explore one-vs-rest classification models and sequence prediction models for multi-label classification tasks. The system ranks first in Bulgarian and second in Russian and Portuguese. In Subtask 3 (Narrative Extraction), we build our system with data augmentation, supervised fine-tuning, and preference-based reinforcement learning. This system achieves the first place in Bulgarian, Russian and Hindi and the second place in Portuguese.

We examine how embedding bias affects hate speech detection by evaluating two debiasing methods—hard-debiasing and soft-debiasing. We analyze stereotype and sentiment associations within the embedding space and assess whether debiased models reduce censorship of marginalized authors while improving detection of hate speech targeting these groups. Our findings highlight how embedding bias propagates into downstream tasks and demonstrates how well different embedding bias metrics can predict bias in hate speech detection.

Recent studies constructing direct interactions between the claim and each single user response (a comment or a relevant article) to capture evidence have shown remarkable success in interpretable claim verification. Owing to different single responses convey different cognition of individual users (i.e., audiences), the captured evidence belongs to the perspective of individual cognition. However, individuals’ cognition of social things is not always able to truly reflect the objective. There may be one-sided or biased semantics in their opinions on a claim. The captured evidence correspondingly contains some unobjective and biased evidence fragments, deteriorating task performance. In this paper, we propose a Dual-view model based on the views of Collective and Individual Cognition (CICD) for interpretable claim verification. From the view of the collective cognition, we not only capture the word-level semantics based on individual users, but also focus on sentence-level semantics (i.e., the overall responses) among all users and adjust the proportion between them to generate global evidence. From the view of individual cognition, we select the top-k articles with high degree of difference and interact with the claim to explore the local key evidence fragments. To weaken the bias of individual cognition-view evidence, we devise inconsistent loss to suppress the divergence between global and local evidence for strengthening the consistent shared evidence between the both. Experiments on three benchmark datasets confirm that CICD achieves state-of-the-art performance.

Recently, neural networks based on multi-task learning have achieved promising performance on fake news detection, which focuses on learning shared features among tasks as complementarity features to serve different tasks. However, in most of the existing approaches, the shared features are completely assigned to different tasks without selection, which may lead to some useless and even adverse features integrated into specific tasks. In this paper, we design a sifted multi-task learning method with a selected sharing layer for fake news detection. The selected sharing layer adopts gate mechanism and attention mechanism to filter and select shared feature flows between tasks. Experiments on two public and widely used competition datasets, i.e. RumourEval and PHEME, demonstrate that our proposed method achieves the state-of-the-art performance and boosts the F1-score by more than 0.87%, 1.31%, respectively.