SCNet: Hate Speech Detection Leveraging Syntactic Information and Conversational Contexts

The rapid growth of social networks has facilitated the spread of hate speech. In recent years, hate speech has become more covert, often disguising its discriminatory intent through euphemisms, sarcasm, and code language. Most existing studies rely on pre-trained language models to generate text embeddings but often overlook syntactic structures and long-distance dependencies. Furthermore, analyzing individual utterances often fails to reveal implicit hate speech tendencies. To address these challenges, we propose SCNet, a novel method that integrates both internal Syntactic information and external Conversational contexts for hate speech detection. First, we represent the syntactic structure within each utterance by Parse Tree Modeling, capturing non-consecutive and long-distance dependencies between words. Then, by constructing a Comment-Repost Graph and a Temporal Comment Sequence, we model the conversation structure on social platforms, capturing cross-level dependencies and long-distance external semantics between utterances. Experimental results demonstrate that SCNet significantly improves hate speech detection performance across two datasets. This study offers a novel approach and valuable insights for advancing hate speech detection.