Hyperbolic Graph Convolutional Networks for Aspect-Based Sentiment Analysis

Abstract

Aspect-based sentiment analysis is a fine-grained sentiment analysis task that aims to predict the sentiment polarity of a specific aspect. Recent work adopts graph convolutional networks over dependency trees to capture the syntactic connections of aspects and opinion words while introducing the BiAffine to jointly refine syntax structures and semantic correlations. However, in the Euclidean space, the neural network models can’t well capture the syntactic connections of aspects and opinion words due to the inaccurate dependency trees representation, and the original structures and correlations are affected due to the BiAffine exchange method. Fortunately, dependency trees can be represented well since hyperbolic space can be viewed as continuous simulations of trees, so we propose a hyperbolic graph convolutional networks (HyperGCN) model to handle these challenges. We employ hyperbolic graph convolution with the dependency tree to model syntactic connections between aspects and opinion words, additionally, we also capture the semantic correlations with a hyperbolic graph convolutional network incorporating self-attention mechanism. Particularly, to exchange the relevant features without original syntax structures and semantic correlations being affected, we leverage an attention mechanism with residual structure to exchange relevant features of syntactic and semantic information. The experimental results on three datasets verify the effectiveness of our model.