Abstract: Emotion analysis of Internet public opinion focuses on mining deep emotional information in do-main-specific texts, which is of great significance for timely assessment and resolving public opinion risks. Most of the previous work relied on basic emotion knowledge such as emotion symbols and lexical properties to construct semantic features of emotions, ignoring fine-grained linguistic emotion expressions such as holders and cues of emotions in texts. Therefore, according to the characteristics of Internet public opinion data during COVID-19, a synchronous dual-channel recurrent network is introduced to extract fine-grained emotional information. On this basis, an auxiliary sentence construction method and an emotional expression perceptive network based on BERT (BERT-EEP) are proposed, and fine-grained emotion information is used to assist label classification. The dependence relationship between auxiliary information and context is learned through multi-head attention mechanism and bidirectional gated loop unit to realize emotion analysis. To evaluate the validity of the proposed method, a COVID-19 Chinese mood dataset with fine-grained representation is constructed. Experimental results show that the proposed method can effectively perceive the fine-grained emotion information and achieve significant performance on the emotion classification task.

Key words: public opinion analysis, emotion classification, fine-grained, emotional expression perception