Enhancing aspect-based sentiment analysis using data augmentation based on back-translation

Aspect-based sentiment analysis (ABSA) identifies mentioned aspects and predicts their associated sentiments in sentences. With the rapid growth of users' online activities, ABSA, as a means of automatically interpreting text, shows significant importance. Aspect terms and opinions are dissimilar for different topics; hence, providing more labeled data related to the domain might be required to achieve better performance. Labeling massive amounts of data is expensive and time-consuming, but by using data augmentation, enhancing the performance is possible without collecting new labeled data. In this paper, we present a hybrid data augmentation method to extend the original data and increase ABSA performance. Back-translation has proved helpful in other areas of NLP as a paraphrasing method to augment text, but because of the data structure of ABSA, it has not been used at its full potential in this field yet. We utilize Special Character Insertion (SCI) in back-translation to address this compatibility issue and generate synthetic augmented sentences. By doing so, the output sentences will preserve meaning and sentiments toward the aspect terms, and their location in the sentence will be restored. Using Random Entity Replacement (RER), we can make the back-translated sentence even more diverse to make the model generalize better on the limited data. RER is used to replace named entities with the lowest chance of getting replaced by the back-translator. Different ABSA models and two benchmark datasets, SemEval 2014 in Restaurant and Laptop domains, are used to evaluate our approach, and five different languages as the middle language for back-translation are investigated. Results show that using German as a middle language, our approach on average can increase accuracy and f1 score by 0.78 and 1.75 percent compared to the original dataset.