Aspect sentiment triplet extraction based on data augmentation and task feedback

Aspect sentiment triplet extraction (ASTE), which focuses on mining the triplets (aspect, opinion, sentiment), is a complex and integrated subtask in aspect-based sentiment analysis. It is widely used in market research, product design and promotion, online comment analysis, and so on. Although significant progress has been achieved in existing methods, several challenges remain, such as data scarcity and the separation of span extraction and sentiment classification. Therefore, this paper adds data augmentation and task feedback based on the bidirectional machine reading comprehension model. Before training the model, the data augmentation module applies mask prediction and mark replacement to enrich the data. Span extraction and sentiment classification are two tasks during ASTE. We adopt the direct span extraction method with one classifier to avoid the error accumulation caused by multiple classifiers and to improve the adaptive ability between different datasets. In addition, we fuse the text features derived from the above two tasks for sentiment classification. Based on the feature fusion, the task feedback module is established to alleviate the task separation. Extensive experiments verify the effectiveness of our method. The code is available at https://github.com/zhenzhen313/BMRC-with-DA-and-TF.