A Span-based Enhanced Bidirectional Extraction Framework for Multi-word Aspect Sentiment Triplets

Abstract

Aspect Sentiment Triplet Extraction (ASTE), aiming to jointly identify all the aspect terms, opinion terms and their corresponding sentiment polarities simultaneously, is a most recent fine-grained sentiment analysis subtask. There are two main categories of ASTE methods, named pipeline approaches and tagging-based joint extraction approaches. The former suffers from error propagation and the latter fail to handle one-to-many and many-to-one problems in triples. In order to address these issues, we propose a span-based enhanced bidirectional extraction framework. The framework utilizes all possible candidate spans as input and adopts syntactic dependency tree to fully explore sentence features. The proposed model extracts triples bilaterally from two directions, to handle multi-word triples and complex correspondence problems between aspects and opinions. In our framework, dual-channel pruning strategy is introduced to choose the correct spans. Meanwhile, bidirectional transformer-based decoders are proposed to model the association among spans and then to extract corresponding triplets. Experiments on four benchmark datasets indicate that our framework reveals significant performance improvement compared to the current state-of-the-art model, especially in predicting triplets with multi-word targets or opinions.