Syntactic and semantic dual-enhanced bidirectional network for aspect sentiment triplet extraction

Span-level method achieves competitive results in Aspect Sentiment Triplet Extraction (ASTE) by enumerating all possible spans. However, previous span-level methods fail to exploit syntactic information to identify the correspondence between aspect terms and opinion terms, which makes the extracted triplets inaccurate. In this paper, we propose a syntactic and semantic dual-enhanced bidirectional network (SSBN) for ASTE task. By constructing word dependencies as a graph and embedding them into features to capture syntactic information more effectively in bidirectional network. Furthermore, we design a pruning strategy that uses part-of-speech information to alleviate the problem of identifying potential aspects and opinions from a large number of spans. We conduct extensive experiments on four benchmark datasets, and the experimental results demonstrate the effectiveness of the SSBN model.