Abstract
Weibo, one of the most widely used social media platforms in China, sees a vast number of users expressing their opinions and emotional tendencies. Conducting sentiment analysis on Weibo posts using natural language processing techniques is crucial for market research and public opinion observation, holding significant commercial and societal importance. However, Chinese expression is highly diverse, making sentiment polarity harder to discern. Traditional sentiment classification algorithms often struggle with insufficient semantic feature extraction and coarse-grained in sentiment classification for Chinese texts. To address these challenges, this paper proposes a Chinese sentiment multi-classification method based on ERNIE-MCBMA. The proposed model extracts the parallel local dependency features between words through the multi-channel CNN convolutional layer and then uses the collaborative architecture of bidirectional LSTM and multi-head attention mechanism to realize context-sensitive feature recalibration. Finally, the shallow syntactic features and deep semantic representation were fused through the cross-layer feature fusion layer to realize the complementary enhancement of more fine-grained semantic information. The SMP2020-EWECT public dataset is used to categorize texts into six classes: neutral, happy, angry, sad, fear, and surprise. Various comparative experiments were conducted on the dataset. The experimental results show that the ERNIE-MCBMA achieves an accuracy of 78.26% and an F1-score of 78.45% for the 6-class classification task, outperforming other baseline models.