Simplifying Aspect-Sentiment Quadruple Prediction with Cartesian Product Operation

Abstract

Aspect sentiment quad prediction (ASQP) is an emerging subtask of aspect-based sentiment analysis, which seeks to predict the sentiment quadruplets of aspect terms, aspect categories, associated sentiment polarities, and corresponding opinion items in one shot. Recent studies employ text generation models to accomplish this task. However, there are still two problems, how to effectively reduce the ASQP task’s high complexity, and the possibility that the generative model may predict explicit terms that do not exist in text sentences. In order to fill the gap, this paper proposes a novel text generation model Cartesian-ASQP based on the Transformer architecture. Specifically, this paper simplifies the aspect-based sentiment quad prediction task to a sentiment triple extraction task by performing a Cartesian product operation on the aspect categories and sentiment polarity sets. For sentiment quadruplet text sentences containing pronouns as implicit terms, we present an implicit term processing strategy by semantically mapping these terms back to pronouns. On the output side, for the situation when the explicit aspect/opinion words predicted by the model are absent from input sentences, this paper introduces a two-stage term correction strategy to solve the problem. Experimental results on two publicly available datasets demonstrate that our proposed model outperforms various baseline methods and achieves outperform performance. This work also validates that our proposed model can effectively handle the task of aspect-based sentiment quad prediction with a large number of implicit aspect and opinion terms.