Aspect-Level Sentiment Classification Based on Graph Attention Network with BERT

Abstract

With accelerated evolution of the internet, people can express their sentiments towards organizations, politics, products, events, etc. Analyzing these sentiments becomes very beneficial for businesses, government and individuals. To some extent, text sentiment analysis involves Internet content security. Aspect level sentiment classification aims to recognize the sentiment expressed towards a special target given a context sentence, which is more fine-grained than sentence level or document level sentiment analysis. Most previous works concentrate on the semantic information between the contexts and the aspect terms, but they ignore the structural information of the sentences. In this paper, we propose a new method based on Graph Attention Network (GAT) to deal with the complex connections among words in sentences through structural features. A lot of experiments are conducted on two datasets: Laptop and Restaurant from SemEval2014, the results show that our model outperforms other previous works, which confirms the effectiveness of taking sentence structural information into account.