System Report for CCL25-Eval Task 10: SRAG-MAV for Fine-Grained Chinese Hate Speech Recognition

Effective hate speech detection has become a critical focus in Natural Language Processing (NLP), aiming to mitigate these negative impacts (Davidson et al., 2017; Waseem and Hovy, 2016). Moreover, ensuring the fairness of detection models to avoid potential biases is essential for their practical deployment (Sap et al., 2019). Traditional methods often rely on binary classification to identify hateful content (Fortuna and Nunes, 2018), but these approaches lack the granularity to capture the internal structure of hate speech, limiting their interpretability and utility for downstream applications (Yin and Zubiaga, 2021). Consequently, Fine-Grained Chinese Hate Speech Recognition (FGCHSR), which extracts structured information such as specific targets or types of hate, has gained increasing attention (Basile et al., 2019; Mathew et al., 2021; Ren et al., 2021). CCL25-Eval Task 10 focuses on extracting quadruplets (Target, Argument, Targeted Group, Hateful) from Chinese social media texts. This task is particularly challenging due to the subtle, context-dependent nature of Chinese hate speech (Pavlopoulos et al., 2020), the interdependence of quadruplet elements, and the limited availability of high-quality annotated data (Yin and Zubiaga, 2021). The ST A TE ToxiCN study (Bai et al., 2025) highlights these difficulties, showing that even the most advanced models like GPT -4o achieve an Average Score of only 15.63, while fine-tuned open-source models like Qwen2.5-7B reach 35.365, but still require further optimization. To address these challenges, we propose a novel SRAG-MA V framework synergistically combining Task Reformulation (TR), Self-Retrieval-Augmented Generation (SRAG), and Multi-Round Accumulative V oting (MA V). Our approach simplifies quadruplet extraction into triplet extraction, enhances contextual understanding through dynamic retrieval inspired by Retrieval-Augmented Generation (RAG) Corresponding author: Jing Li ( jingli.phd@hotmail.com).