Abstract: Aiming at the problems of insufficient local information capture and multiple opinion words confusion in aspect level sentiment analysis, this paper proposes an aspect level sentiment analysis model based on word co-occurrence. In this model, aspect level sentiment analysis is regarded as a sentence pair task, and the node information including context and aspect word interaction attention is obtained by BERT. At the same time, independent word co-occurrence graph is constructed for each data sample, and gating graph neural network is used to update nodes to enhance the fusion of information near aspect words and reduce the interference of irrelevant opinion words. Then, the global information is further fused in the self-attention layer. Finally, the aspect word nodes are sent to the nonlinear layer to obtain the classification results. Compared with six baseline models, the experimental results show that the model can effectively improve the accuracy of aspect level sentiment analysis. 

Key words: aspect-level sentiment analysis, gated graph neural network, word co-occurrence graph, self-attention, BERT