Span-based dual-decoder framework for aspect sentiment triplet extraction

Abstract

Aspect sentiment triplet extraction (ASTE) aims to extract aspects from review sentences along with their corresponding opinions and sentiments to form opinion triplets. Since each factor within a sentiment triplet could be a single word or a phrase, defining and implementing the span-level features for this span-level task is critical and challenging. However, prior works typically formulate the ASTE task as a sequence tagging problem and address it with token-level models, limiting the extraction performances of long entities and suffering from cascading errors due to sequential decoding. Although some methods have enumerated all possible spans as input, they fail to explicitly build the interaction among the potential triplets and semantic information within the sentence explicitly. To address these problems, we propose a span-based joint training framework, where each potential entity is represented as an independent span and sentiment polarity is classified by using the corresponding independent span representations. Specifically, we design two different transformer-based decoders to extract the aspects and their corresponding opinions, respectively. Those decoders utilize multiple multi-head attention mechanisms to model the associations among the spans and the semantic information between the spans and the sentences. To verify the effectiveness of our approach, we conduct extensive experiments on four benchmark datasets. The experimental results demonstrate that our proposed method significantly outperforms existing state-of-art methods.