Domain-level Pairwise Semantic Interaction for Aspect-Based Sentiment Classification

TL;DR

This paper introduces a Pairwise Semantic Interaction module that leverages domain-level relationships between sentences to improve aspect-based sentiment classification, especially addressing class-imbalance issues.

Contribution

It proposes a novel plug-and-play PSI module that models pairwise sentence interactions to enhance semantic representations and mitigate class-imbalance in ABSC.

Findings

PSI outperforms several state-of-the-art baselines.

PSI significantly alleviates class-imbalance problems.

Experimental results on four datasets validate effectiveness.

Abstract

Aspect-based sentiment classification (ABSC) is a very challenging subtask of sentiment analysis (SA) and suffers badly from the class-imbalance. Existing methods only process sentences independently, without considering the domain-level relationship between sentences, and fail to provide effective solutions to the problem of class-imbalance. From an intuitive point of view, sentences in the same domain often have high-level semantic connections. The interaction of their high-level semantic features can force the model to produce better semantic representations, and find the similarities and nuances between sentences better. Driven by this idea, we propose a plug-and-play Pairwise Semantic Interaction (PSI) module, which takes pairwise sentences as input, and obtains interactive information by learning the semantic vectors of the two sentences. Subsequently, different gates are generated to effectively highlight the key semantic features of each sentence. Finally, the adversarial interaction between the vectors is used to make the semantic representation of two sentences more distinguishable. Experimental results on four ABSC datasets show that, in most cases, PSI is superior to many competitive state-of-the-art baselines and can significantly alleviate the problem of class-imbalance.