Interactive Attention Graph Convolution Networks for Aspect-level Sentiment Classification

Abstract

Different from coarse-grained sentiment analysis, fine-grained sentiment analysis can identify the sentiment orientation of a given aspect in the context. Although neural network model with attention mechanism perform well, most of them focus only on semantic analysis of annotations, with little consideration of syntactic constraints between aspects and context. How to efficiently use syntactic dependency information to optimize the representation of contexts is an important issue. An Interactive Attention Graph Convolutional Network (IAGCN) was constructed by us. In order to use syntactic information, the network first extracts syntactic information on the syntactic dependency tree with the help of the graph convolutional network. Then, through interactive learning, various aspects and contextual representations are generated. In this design, the aspects and context of syntactic information are fully learned. The experimental results on 5 data sets verify the advantages of the model.