BHF: BERT-based Hierarchical Attention Fusion Network for Cyberbullying Remarks Detection
Ziyang Feng, Jintao Su, Junkuo Cao · 2022
Cyberbullying remarks based on social media platforms have seriously endangered the physical and mental health of netizens. In order to purify the network environment and create a healthy atmosphere, it needs to be identified. At present, most of the existing studies only analyze the overall features of the text, while ignoring the semantic representation of words at local locations. In order to mine the richer semantic features of text, this paper proposes a BHF model, which uses the BERT pre-training model to extract global semantic feature information, and then uses the Hierarchical Attention Network (HAN) to calculate the two-dimensional space of words and short sentences. Local semantic feature information is obtained by fusing two layers of semantic feature information to obtain three-dimensional semantic feature representation of words, short sentences and full text. The experimental results show that based on the BHF model, the semantic features of the text can be better extracted on the three types of problems of cyberbullying detection, and the accuracy of cyberbullying remarks detection can be greatly improved.