Incorporating Hybrid Pooling and Attention Mechanisms for Chinese Text Sentiment Analysis

An improved sentiment analysis model is designed to address the problems of increasingly spoken and fragmented web texts and the difficulty of extracting textual sentiment features. At first, a pre-trained model of BERT-wwm(Bidirectional Encoder Representations from Transformers-Whole Word Masking) is used to generate textual sentence vector features; second, textual local semantic features are extracted by a hybrid pool of convolutional neural networks and contextual textual semantic features are extracted by a bidirectional recursive gating network that includes an attention mechanism. Finally, the multilevel text features are fused and classified using a classifier. The experimental results show that the F1 values on the microblog text dataset weibo_senti_100k and the COVID-19 pandemic microblog comment dataset are 98.16% and 92.55%,which are better than the benchmark model.