Modelling Context with Graph Convolutional Networks for Aspect-based Sentiment Analysis

Aspect-based sentiment analysis is a fine-grained natural language processing task that aims to predict a specific target's sentiment polarity in its context. Existing researches mainly focus on the exploration of the interaction between the sentiment polarity of aspects and contexts. Models based on the self-attention mechanism can fully explore the syntactic structure of sentences. In contrast, models based on a convolutional neural network have the ability to make aspects and the semantics of contextual words alignment. These methods all have some limitations; that is, they lack the ability to make full use of syntactic information and long-range word dependencies to carry out relevant syntactic constraints while associating the target's sentiment with the local context. And they are not able to handle affective ambivalence in text. In this paper, we propose a stacked ensemble method for predicting the sentiment polarity by combining a local context embedding and a global graph convolutional network. It uses a Graph Convolutional Network (GCN) to supplement local information to improve the accuracy of the aspect sentiment classifier with revealing multi-level sentiments. Experimental results on three commonly used datasets show that our approach outperforms the state-of-the-art models in the Semeval-2014 dataset.