A multi-task learning network using shared BERT models for aspect-based sentiment analysis

Abstract Aspect-based sentiment analysis (ABSA) aims to predict the sentiment polarity of specific aspect words occurring in a text. ABSA includes aspect-category sentiment analysis (ACSA) and aspect-target sentiment analysis (ATSA). There have been many previous studies addressing both tasks through RNNs and other neural models. With BERT's remarkable performance on NLP tasks, several studies have enhanced aspect word extraction to solve ATSA by building new BERT-based models. But such an approach is not directly applicable to the ACSA task, because aspect words in ACSA are often not explicitly present in the text, so aspect word extraction becomes more difficult. In this paper, we propose a multi-task learning (MTL) approach to solve these problems. Our approach is based on a shared BERT model to construct a multi-task learning network, which is trained by strongly and weakly related tasks. We also use a multi-head self-attention layer to replace a linear layer in traditional multi-task learning networks, to enhance the ability to capture global semantics. We also propose a new fine-tuning strategy that can better improve the performance of the model. Experiments were conducted on four datasets from the ATSA task and four datasets from the ACSA task: laptop, restaurant, restaurant-2014, and restaurant-large from SemEval-2014. Our experimental results show that: In the ACSA task, our model outperformed all the baseline models, achieving the current state-of-the-art performance on the multiple datasets. For the ATSA task, our model performs close to the state-of-the-art, with much simpler architecture.