Span-Based Attention Decoder Framework For Aspect Sentiment Triplet Extraction

Aspect sentiment triplet extraction (ASTE) aims to extract aspect terms, opinion terms and sentiment polarity from the comment sentence. Recent ASTE models rely on the word-level interaction without considering the context representation in sentiment prediction, which limits the performances of sentiment prediction and triplet extraction. We propose a span-based attention decoder model (SBAD) by using the span-level interaction between aspect and opinion when predicting pairs sentiment. SBAD matches the selected aspect term with the opinion term one by one, and the pairs are considered to predict sentiment polarity. To further capture sentiment features from context representation, we design a two-layer multi-head attention decoder to decode the interaction among aspect, opinion and context representation. In the first layer, each pair relationship between aspect and opinion is captured. In the second layer, the sentiment features of the context representation that the pair concern is further captured. To verify the effectiveness of our model, we conduct several experiments on ASTE datasets. The results show that the proposed model significantly outperforms the strong baseline models on four ASTE datasets.