Cross-domain aspect-based sentiment analysis using domain adversarial training

Abstract

Over the last decades, the increasing popularity of the Web came together with an extremely large volume of reviews on products and services useful for both companies and customers to adjust their behaviour with respect to the expressed opinions. Given this growth, Aspect-Based Sentiment Analysis (ABSA) has turned out to be an important tool required to understand people’s preferences. However, despite the large volume of data, the lack of data annotations restricts the supervised ABSA analysis to only a limited number of domains. To tackle this problem a transfer learning strategy is implemented by extending the state-of-the-art LCR-Rot-hop++ model for ABSA with the methodology of Domain Adversarial Training (DAT). The output is a cross-domain deep learning structure, called DAT-LCR-Rot-hop++. The major advantage of DAT-LCR-Rot-hop++ is the fact that it does not require any labeled target domain data. The results are obtained for six different domain combinations with testing accuracies ranging from 35% up until 74%, showing both the limitations and benefits of this approach. Once DAT-LCR-Rot-hop++ is able to find the similarities between domains, it produces good results. However, if the domains are too distant, it is not capable of generating domain-invariant features. This result is amplified by our additional analysis to add the neutral aspects to the positive or negative class. The performance of DAT-LCR-Rot-hop++ is very dependent on the similarity between distributions of source and target domain and the presence of a dominant sentiment class in the training set.