面向中文细粒度仇恨识别的两阶段多轮问答框架
A Two-Stage Multi-Round Question-Answering Framework for Fine-Grained Chinese Hate Speech Recognition
李文煜 1张雪 1陈钰枫1
作者信息
- 1. 北京交通大学计算机科学与技术学院，北京 100044
- 折叠
摘要
仇恨言论的精准识别是网络内容安全治理的重要环节，对维护社会舆论环境与公共秩序具有现实意义。然而，现有中文细粒度仇恨识别方法在片段级要素抽取能力不足，难以准确定位文本中的攻击对象、论点与仇恨群体，对隐晦表达与复杂语义结构的处理亦存在明显局限。为解决上述问题，该文提出面向中文细粒度仇恨识别的两阶段多轮问答框架，构建多轮提示结构并引入自检索增强机制，通过多轮累积投票提升输出稳定性与一致性。实验在 STATE-ToxiCN 数据集上表明，该方法所有细粒度元素匹配得分超过当前CCL2025评测基线与领先方案，在四元组层面软硬匹配F1平均值达38.85。结果验证了多轮提示与检索增强在提高模型鲁棒性和复杂语义识别方面的有效性，为中文细粒度仇恨识别及多层次信息抽取提供可行方案。
Abstract
Accurate identification of hate speech is a crucial component of online content safety governance and is of practical significance for maintaining a healthy public opinion environment and social order. However, existing Chinese fine-grained hate speech identification methods remain weak in segment-level element extraction: they struggle to precisely locate attack targets, arguments, and hateful groups in text, and show clear limitations in handling implicit expressions and complex semantic structures. To address these issues, this paper proposes a two-stage multi-round question-answering framework for fine-grained Chinese hate speech recognition. The framework designs multi-round prompting and incorporates a self-retrieval augmented mechanism, while using multi-round accumulated voting to improve the stability and consistency of model outputs. Experiments on the STATE-ToxiCN dataset show that the proposed method outperforms the current CCL2025 evaluation baseline and leading approaches across all fine-grained element matching scores, achieving an average soft/hard matching F1 of 38.85 at the quadruple level. These results demonstrate the effectiveness of multi-round prompting and retrieval augmentation in enhancing model robustness and recognizing complex semantics, providing a feasible solution for fine-grained Chinese hate speech identification and multi-level information extraction.关键词
大语言模型/提示工程/仇恨言论识别Key words
large language model/prompt engineering/hate speech recognition引用本文复制引用
李文煜,张雪,陈钰枫.面向中文细粒度仇恨识别的两阶段多轮问答框架[EB/OL].(2026-01-26)[2026-01-29].http://www.paper.edu.cn/releasepaper/content/202601-55.学科分类
语言学/汉语
评论