Using User’s Expression Propensity for Sarcasm Detection Based on Sequential Three-Way Decision

Abstract

AbstractSarcasm detection is mainly to distinguish whether the target comment is sarcasm that can help identify the actual sentiment. The previous sarcasm detection mainly focused on text features using vocabulary, grammar, and semantics. But users’ expression propensity is ignored which is helpful to distinguish some comments with uncertain sarcasm polarity in sarcasm detection. However, how to use the user’s expression propensity for sarcasm detection effectively is a challenge. Based on the ideas of granular computing and three-way decisions, we propose a sarcasm detection model based on the sequential three-way decision (S3WD) to integrate text features and users’ expression propensity. The S3WD divides the comments into the sarcasm (SAR) region, non-sarcasm (NSAR) region, and boundary region (BND), and then gradually divides the uncertain BND region into a clear SAR region and NSAR region. We firstly construct a sequential structure through analysis sentiment of comments’ chunks. Second, text features and users’ expression propensity are fed into different sequential layers for fusion that can guide the comment classification more effectively. Finally, contextual information is further applied to consider sentiment context during sarcasm detection. The experimental results on a large Reddit corpus show that our model improves sarcasm classification performance effectively.KeywordsSarcasm detectionSequential three-way decisionUser’s expression propensity