Comparative Analysis on Aspect-based Sentiment using BERT

Abstract

Aspect-based Sentiment Analysis (ABSA) is a complex model within the domain of Sentiment Analysis (SA) tasks which deals with classifying the sentiments related to particular aspects (or targets) in the given text. ABSA task has gained popularity due to its various sub-tasks related to the aspect-based sentiment analysis task. This work provides a comparative study of various approaches used to solve the ABSA task using the BERT technique. The selected approaches include a fine-tuned BERT model, adversarial training using BERT (Bidirectional Encoder Representations from Transformers) and the incorporation of disentangled attention in BERT or the DeBERTa for the ABSA task. One of the challenges faced during implementation of the ABSA task is that it requires an in-depth understanding about the language. Experiment results indicate that the approach, which uses the fine-tuned BERT model yields the best mean F1 score of 85.65 and the best mean accuracy score of 85.98 is yielded by the DeBERTa model.