A Joint Model for Aspect-Category Sentiment Analysis with TextGCN and Bi-GRU

Abstract

In the age of Internet, online customers express their opinions on products by posting reviews. It is critical to do sentiment analysis on customers’ review data to help subsequent customers make their purchasing decisions and guide companies to improve their products. Aspect-Category Sentiment Analysis(ACSA) is a subtask of sentiment analysis, which aims to detect the aspect categories mentioned in the text and identify their corresponding sentiment polarities. Most of the previous methods regard Aspect Category Detection(ACD) and AspectCategory Sentiment Classification(ACSC) as two separate tasks, which could lead to error propagation so as to lack practicality. However, as far as we know, joint modeling of aspects and polarities has not yet received widespread attention. The only few existing joint models do not work well with representing the relation of aspect categories and its corresponding sentiment words. To address these problems, we propose a novel model (TG-GRU) based on the combination of GCN and Bi-GRU for detecting aspect categories and corresponding sentiment polarities simultaneously. Experiments are conducted on one Chinese dataset and two English datasets. Experimental results show that our model outperforms the state-of-the-art methods of joint models.