BERT-based multi-task learning for aspect-based sentiment analysis

Abstract

The Aspect Based Sentiment Analysis (ABSA) systems aims to extract the aspect terms (e.g., pizza, staff member), Opinion terms (e.g., good, delicious), and their polarities (e.g., Positive, Negative, and Neutral), which can help the customers and companies to identify product weaknesses. By solving these product weaknesses, companies can enhance customer satisfaction, increase sales, and boost revenues. There are several approaches to perform the ABSA tasks, such as classification, clustering, and association rule mining. In this research we have used a neural network-based classification approach. The most prominent neural network-based methods to perform ABSA tasks include BERT-based approaches, such as BERT-PT and BAT. These approaches build separate models to complete each ABSA subtasks, such as aspect term extraction (e.g., pizza, staff member) and aspect sentiment classification. Furthermore, both approaches use different training algorithms, such as Post-Training and Adversarial Training. Moreover, they do not consider the subtask of Opinion Term Extraction. This thesis proposes a new system for ABSA, called BERT-ABSA, which uses MultiTask Learning (MTL) approach and differentiates from these previous approaches by solving all three tasks such as aspect terms, opinion terms extraction, and aspect term related sentiment detection simultaneously by taking advantage of similarities between tasks and enhancing the model’s accuracy as well as reduce the training time. To evaluate our model’s performance, we have used the SemEval-14 task 4 restaurant datasets. Our model outperforms previous models in several ABOM tasks, and the experimental results support its validity