Multi-Tasking for Aspect-Based Sentiment Analysis via Constructing Auxiliary Self-Supervision ACOP Task

Abstract

Aspect-based sentiment analysis (ABSA) is a fine-grained task of sentiment analysis. In aspect-based sentiment analysis, the semantic understanding of sentences is critical. The previous approach had to design modules to improve the understanding of sentence semantics. However, We adopt a multi-task approach to learn sentence semantics to avoid design modules. Thus we propose the Aspect Word and Context Order Prediction Task (ACOP) as an auxiliary task. We implement the ACOP task with the global and local way for aspect-based sentiment analysis and adopt the self-supervised method to train our model. Our model improves accuracy and F1 values over the best baseline model on the Rest14 dataset by 1.96% and 1.76%, on the Lap14 dataset by 0.24% and 0.24%, and on the Twitter dataset by 1.45% and 2.03%. Our experiments on three public datasets demonstrate that our approach is effective and that using a multi-task approach is a good choice instead of designing corresponding modules to extract semantic features.