Deliberate Multi Parallel Graph Convolution Network for Aspect-Based Sentiment Analysis

Aspect-based sentiment analysis (ABSA) is a fine-grained sentiment classification task, which needs to detect the sentiment polarity towards a given aspect. Recently, graph neural networks over the dependency tree have been widely applied for aspect-based sentiment analysis. Most existing works generally focus on the extraction of syntactic information, which ignores the impact of different dependencies. In this work, we propose a novel structure named deliberate multi-parallel graph convolutional network (DMP-GCN) that not only extracts the semantic information of the sentence itself but also captures the syntactic information according to different dependencies. To be specific, the proposed DMP-GCN aggregates two blocks, a multi-head attention semantic extraction block (MAS-Block) and a multi-parallel syntactic extraction block (MPS-Block). We construct a set of dependency graphs in MPS-Block to enhance the representations of different dependencies and generate an attention matrix to capture semantic information in MAS-Block. Experimental results on multiple public benchmark datasets illustrate that our proposed model achieves better results compared to other models.