Today, writing an online review is easier than ever, and these reviews significantly influence the reputation of businesses. Customers routinely rely on them when assessing a company’s reputation and deciding whether to purchase a product or use a service. Therefore, Aspect Sentiment Classification (ASC), which determines the sentiment for a given aspect category, has attracted considerable attention. One of the main bottlenecks in ASC research is the limited availability of annotated data. To address this issue, in this paper, we propose the use of data augmentation techniques. We investigate Back Translation (BT), Easy Data Augmentation (EDA), Keyboard Augmentation (KA), and mixup, with the aim of improving the robustness and overall performance of ASC models. As ASC models, we employ ASC-BERT-Pair and ASC-RoBERTa-Pair, applied at the document level. Using the SemEval 2016 reviews dataset, we demonstrate that out of the examined data augmentation methods, Keyboard Augmentation is the most effective, and ASC-RoBERTa-Pair outperforms ASC-BERT-Pair.
Duits et al. (Thu,) studied this question.