Fine grained sentiment analysis on microblogs based on graph convolution and self attention graph pooling

Microblog is one of the most popular social media platforms in China. Fine grained sentiment analysis of Chinese microblog comments has attracted much attention. Graph Convolutional Neural Network (GCN) has been broadly used in sentiment analysis but still suffers from emotion misclassification due to the complexity and diversity of Chinese microblogs’ syntax structures. To address the issue, we propose a graph pooling method based on self-attention mechanism, namely, AGMPool. The AGMPool pooling method uses graph convolution to calculate its attention score for each graph node and then to filter out nodes with excessive useless information in the graph topology according to these scores, which effectively improves the performance of fine-grained sentiment analysis through GCN. In addition, for better understanding of diverse syntax structures of Chinese microblogs, we propose a microblog fine grained sentiment analysis model, namely, LMG-AGMPool, which combines GCN with the AGMPool pooling method and extracts emotional features based on the syntax structures of text and the importance of words in text. The experimental results indicate that the LMG-AGMPool model has better performance than the traditional methods and the deep learning methods in fine grained sentiment analysis.