Opinion Targets and Sentiment Terms Extraction based on Self-Attention

Abstract

Opinion targets and sentiment terms extraction is a key task in aspect-level sentiment analysis. Existing researches have shown that using the dependency structures of reviews helps to complete the task. In this paper, we present a novel opinion targets and sentiment terms extraction model based on self-attention. The proposed model learns the contextual feature of each token in a review through a LSTM network, and a self-attention mechanism is used to directly capture the relations between any two tokens in a review, thus the global dependencies and internal structure of reviews can be modeled better. Experiments results on two benchmark datasets show that the proposed model achieves a better performance than current state-of-the-art in opinion targets and sentiment terms extraction task.