Research on sentiment analysis methods based on aspect word embedding graph convolutional networks

Abstract

Aspect-based sentiment analysis is a fine-grained task in the field of sentiment analysis. Various GCN approaches have recently emerged to work on this, but many approaches ignored the critical role of aspectual word information and the effect of noise. In view of this situation, we propose an aspect-based word embedding graph convolutional network (AWEGCN) model. In order to make good use of the aspect information and distinguish the contextual information that is more important for a particular aspect, the aspect information is embedded in the output of the hidden layer. To reduce the noise effect when multiple aspect words appear in a sentence, after going through the bidirectional graph convolutional network, the aspect information is embedded. A specific contextual representation is computed through an attention mechanism, which is used as the final classification feature. Experiments show that our model achieves impressive performance on five public datasets, and we also apply BERT and XLNet pre-trained models to this task and obtain advanced results that validate the effectiveness of our model.