Graph convolutional network with multiple weight mechanisms for aspect-based sentiment analysis

Aspect-based sentiment analysis (ABSA) aims at determining the sentiment polarity of the given aspect term in a sentence. Recently, graph convolution network (GCN) has been used in the ABSA task and obtained promising results. Despite the proliferation of the methods and their success, prevailing models based on GCN lack a powerful constraint mechanism for the message passing to aspect terms, introducing heavy noise during graph convolution. Further, they simply average the subword vectors from BERT to form word-level embeddings, failing to fully exploit the potentials of BERT. To overcome these downsides, a graph convolutional network with multiple weight mechanisms is proposed for aspect-based sentiment analysis in the paper. Specifically, a dynamic weight alignment mechanism is proposed to encourage our model to make full use of BERT. Then an aspect-aware weight mechanism is designed to control message propagation to aspect during graph convolution operation. Afterwards, an aspect-oriented loading layer is presented to further reduce adverse effects of words irrelevant with aspect term. Finally, the multi-head self attention is used to fuse high order semantic and syntax information. Hence, the model can obtain the premium aspect-specific representations for prediction. Experiments demonstrate that the proposed model can achieve state-of-the-art results compared to other models.