RATING-AWARE MACHINE LEARNING FOR SARCASM DETECTION IN TEXT–IMAGE MULTIMODAL REVIEWS

Abstract

Sarcasm remains one of the most challenging aspects of sentiment analysis, as surface-level expressions often contradict the true intent of the reviewer. In online reviews, this challenge is amplified when textual content, visual memes/images, and numerical ratings conflict, misleading both users and recommendation systems. This paper introduces a rating-aware multimodal machine learning framework for sarcasm detection that integrates textual, visual, and rating inconsistency features. Unlike deep learning-based methods, the proposed approach leverages advanced machine learning concepts, including meta-feature projection, contradiction-driven feature extraction, and rating-boosted support vector machines (M³-SVM). The framework explicitly models sentiment–rating discrepancies and projects multimodal features into a contradiction space to enhance discriminative power. Experimental evaluations demonstrate improved accuracy and robustness over baseline classifiers, highlighting the framework’s potential to strengthen trust in review-driven decision-making and improve recommendation reliability.