Domain Adversarial Training for Aspect-Based Sentiment Analysis

The continuously expanding digital possibilities, increasing number of social media platforms, and growing interest of companies in online marketing increase the importance of Aspect-Based Sentiment Analysis (ABSA). ABSA focuses on predicting the sentiment of an aspect in a text. In an ideal scenario, we would have labeled data for every existing domain, but acquiring annotated training data is costly. Transfer learning resolves this issue by building models that can be employed in different domains. The proposed work extends the state-of-the-art LCR-Rot-hop++ model for ABSA with the methodology of Domain Adversarial Training (DAT) in order to create a deep learning adaptable cross-domain structure, called the DAT-LCR-Rot-hop++. The major advantage of the 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 37% up until 77%, showing both the limitations and benefits of this approach. Once DAT is able to find the similarities between domains, it produces good results, but if the domains are too distant, it is not capable of generating domain-invariant features.

Automated summary

With the continuous expansion of the digital domain and the rise of online marketing, aspect-based sentiment analysis has become increasingly important. However, acquiring labeled training data can be costly, making transfer learning an effective solution. This study applies domain adversarial training to the state-of-the-art model and achieves satisfactory results.