Chinese Social Media Offensive Text Detection Method Based on Semantic Homogeneity

At present, offensive information detection on Chinese social media platforms has become an important research direction in the field of text detection. However, the commonly used detection methods still have the following limitations: Dictionary-based methods rely on static offensive word lists, which cannot adapt to the rich and diverse expressions in Chinese; user history-based methods are limited by privacy protection and can only analyze offensive texts that are publicly traceable, which restricts both the depth and breadth of offensive text detection. To address these issues, this paper introduces homogeneity theory into offensive text detection on Chinese social media, and proposes a Chinese offensive text detection model based on homogeneity theory (COTD-HT). This method detects offensive expressions based on internal textual information, identifies the true intention behind different textual expressions, and reduces the dependence on user history information. Specifically, this method first designs a mapping relationship between text and social behavior, and uses similarity and sentence matching algorithms to construct a heterogeneous graph network that captures internal word-sentence relations. Then, offensive label prompts are used to enhance node features, thereby strengthening the expression of offensive features in Chinese texts. Finally, a joint feature extraction method based on GAT and BERT is built for text detection. Experimental results show that using internal textual information for Chinese offensive text detection can reduce the reliance on existing offensive information and achieves good performance. At the same time, the single-label prompting method enhances the model’s preference selection and strengthens the expression of offensive features in text.