Research of Sentiment Analysis Model Fusing Local and Global Information

BERT is widely used pre-trained language model in the field of sentiment analysis. The success of the BERT lies in the powerful role of the self-attention mechanism in Transformer. BERT has shown to be effective at capturing local contextual information in texts, but its ability to capture global information is more limited. In this paper, we propose LCF-GCN model, which uses Graph Convolutional Networks (GCN) to continuously fit adjacency information to capture global features and applies semantic distance calculation to improve local feature extraction based on BERT. The fusion and interaction of local information and global information of domain data are considered comprehensively through the above methods. We conduct comparative experiments on the housing review datasets. Experimental results show that LCF-GCN improves the accuracy of polarity extraction and f1 of aspect extraction task.