Dual-channel and multi-granularity gated graph attention network for aspect-based sentiment analysis

The Aspect-Based Sentiment Analysis(ABSA) aims to determine the sentiment polarity of a specific aspect. Existing approaches use graph attention networks(GAT) to model syntactic information with dependency trees. However, these methods do not consider the noise of the dependency tree and ignore the sentence-level feature. To this end, we propose the Dual-Channel and Multi-Granularity Gated Graph Attention Network(DMGGAT) to jointly consider semantics and syntactic information of multiple granularity features generated by GAT and BERT, in which BERT alleviates the instability of the dependency tree and enhance the semantic information lost in the graph calculation. First, We propose a two-channel structure composed of BERT and GAT, enabling syntactic and semantic information generated by BERT to assist GAT. Furthermore, an aspect-based attention mechanism is used to generate sentence-level features. Finally, a newly designed gated module is introduced to integrate the aspect(fine-Granularity) and sentence-level (coarse-Granularity) features from the two channels to classify jointly. The experimental results show that our model achieves advanced performance compared to the current model on three extensive datasets.