Implicit Offensive Speech Detection Based on Multi-feature Fusion

As social media platforms become increasingly strict in censorship of user posts, aggressive and insulting language has gradually shifted to implicit expressions using homophones, metaphors, and other forms of camouflage. This not only intensifies the satirical attacks of negative posts but also leads to a significant decrease in the effectiveness of models designed to detect offensive speech. This paper aims to achieve high accurate detection of implicit offensive speech. We conducted targeted investigations on the speech characteristics on Weibo, which is one of the largest social media platform in China. Based on the identified features of implicit offensive speech, including semantic, emotional, metaphorical, and fallacy characteristics, this paper constructs a BERT-based Multi-Task learning model named BMA (BERT-Mate-Ambiguity) to accurately detect the implicit offensive speech in the real world. Additionally, this paper establishes a dataset based on posts of Weibo that contain implicit offensive speech and conducts various comparative, robustness, and ablation experiments. The effectiveness of the model is demonstrated by comparing it to existing models that perform well in this field. Finally, this paper discusses some of its limitations and proposes future research work.