SE-GCN: A Syntactic Information Enhanced Model for Aspect-Based Sentiment Analysis

Aspect-based Sentiment Analysis aims to analyze people’s sentiment tendencies towards evaluation targets at the aspect level. Related research in recent years is mainly based on graph convolutional networks, and although much progress has been made, the existing methods focus on utilizing sequence information or syntactic dependency constraints within the text, but without fully utilizing the type of dependency relationships between the aspect terms and the context, and the raw dependency syntactic tree contains noise unrelated to the aspect terms. In this paper, A model for syntactic information-enhanced graph convolutional networks is proposed to address the above problems. The information of dependency relationship types between words is taken as an important feature, and different dependency relationship types are weighted using the attention mechanism. A dependency reconstruction algorithm is also proposed to establish connections between multi-word aspect terms and related viewpoint terms to increase the effective sense field in the convolution process. Experiments on four public datasets demonstrate the effectiveness of the proposed model.