Aspect level sentiment classification with multi-scale information

Abstract

Mining the sentiment polarity of different aspects of products becomes more and more important for companies, as the rapid development of e-commerce. Therefore, aspect-level sentiment analysis technology has attracted widespread attention from researchers. Most of the existing aspect-level sentiment analysis methods are based on feature engineering or neural networks to capture the key points in the sentence, i.e., semantic information. However, most of existing methods mainly pay attention to the impact of a single word in the sentence and simply use individual sentiment words as the basis for the judgement of sentiment polarity, while ignore the importance of the local joint information (i.e., multi-scale information) for a sentence. In order to deal with this problem, we present a novel Aspect-based Multi-scale information Graph Convolutional Network (AMGCN) for aspect-level sentiment analysis, which captures the multi-scale information from the sentence via a series of convolutional neural networks and then incorporate aspect information by a graph convolutional neural network and an attention mechanism. Experimental results over five benchmark datasets demonstrate that the proposed AMGCN can achieve some superior performance in aspect-level sentiment classification when compared to existing methods.