Multi-Dimensional Information Representation for Offensive Language Detection

作者：Peng Yang, Bing Li, Hanlin Zhao, Haojun Huang, Dapeng Wu · 发表于：IEEE Transactions on Emerging Topics in Computational Intelligence · 年份：2025 · DOI：10.1109/tetci.2025.3559094 · 被引用次数：2 · 研究领域：Hate Speech and Cyberbullying Detection

Most offensive language detection techniques for social media are dependent on feature engineering, machine learning, and pretraining models. However, these efforts usually ignore the multi-dimensional analysis of semantic information represented in offensive text, and furthermore, they cannot deeply explore the limited corpora to capture the historical associations of the text. In order to address these problems, in this paper, we propose aMulti-DimensionalInformationRepresentation (MDIR) model for offensive language detection. The proposed model analyzes the deep semantic associations in offensive texts in two dimensions: target and semantics. The former captures the relationship between entities to generate word representation vectors of specific and generalized targets in depth. While the latter combines literal-level information and relevant context, especially user linguistic style and social personality features in historical posts, to extract semantic relations with the sarcastic nature of offensive texts. Moreover, a text vector cross-encoder and a hierarchical fusion detector are conceived to detect the offensive languages accurately by connecting the target dimension and semantic dimension in a mutually reinforcing way. Experimental results show that our MDIR model achieves significant improvements compared to the state-of-the-art baselines over the benchmark dataset. Specifically, the F1 score is 4.18% higher than the current most advanced model, demonstrating the...