Aspect-Based Sentiment Analysis Using The Pre-trained Language Model BERT

Abstract

Sentiment analysis has become popular in both research and business due to the increasing amount of opinionated text generated by Internet users. Sentiment analysis focuses on classifying the overall sentiment of a text, which may not include important information such as different sentiment associated to specific aspects in the text. The more complex task of identifying the sentiment of certain aspects in a text is known as Aspect-Based sentiment analysis (ABSA). This paper show the potential of using the contextual word representations from pre-training language models to solve out-of-domain ABSA by constructing a generic ABSA model using BERT, together with the method of fine-tuning the model to make it learn when aspects are related or unrelated to a text. To our knowledge, no other existing work has been done on out-of-domain ABSA for aspect classification.