BGatCCN: A sentiment analysis network based on BERT utilizing graph attention and dynamic conditional parameter convolution

Sentiment analysis is an essential direction in natural language processing. Previous studies often classified text without fully considering the overall sentiment of the text. To address these issues, we propose a new model that utilizes BERT, a conditional parameter convolutional network, and a graph attention network. Our approach uses hidden state information from the last four layers of BERT. The information is converted into a graph attention matrix through the graph attention layer, which then undergoes conditional parameter convolutions to generate text feature vectors. Max pooling is used to capture the most relevant features for each moment. These features are then combined to obtain global feature information. Finally, we perform sentiment analysis using linear mapping. Through experiments, our model achieves better results on datasets.