Unified Target-based Sentiment Analysis by Dual-pointer Tagging Scheme

Target-based sentiment analysis aims at extracting opinion targets and predicting their sentiment polarities from a given sentence. Previous researches have been trying to joint the two sub-task in a single architecture. However, the target extraction schemes of these approaches are facing various issues, such as ignoring the sentiment consistency of composite targets, or being too sensitive or time-consuming in heuristic search. We propose a model on top of the BERT encoder by designing a dual-pointer binary classifier for target extraction and a joint multi-class classifier for sentiment classification. Experiments on two datasets show the effectiveness of our model.