Enhancing sarcasm detection via multimodal learning: A BiLSTM-attention approach with text and emojis integration

Nasa Zata Dina (Department of Engineering, Faculty of Vocational Studies, Universitas Airlangga, Indonesia)
Moch. Nafkhan Alzamzami (Department of Informatics, Faculty of Intelligent Electrical and Informatics Technology, Institut Teknologi Sepuluh Nopember, Indonesia)

Journal of Soft Computing Exploration
Vol. 7 No. 2 (2026): June 2026

The detection of sarcasm is a difficult task in Natural Language Processing (NLP) because to the presence of implicit meaning and contextual ambiguity. This is particularly problematic in social media, where emojis are used frequently to indicate tone and intent. The study proposes a multimodal deep learning strategy that combines both textual and emoji features, by utilizing a BiLSTM with attention mechanisms. The goal of this method is to improve the performance of sarcasm detection. The model makes advantage of bidirectional contextual learning and preferentially focuses on informative tokens and emojis in order to do more effective work of capturing complex expressions. According to the findings of the experiments, the Text+Emoji model that was proposed achieves an F1-score of 96.44%, an accuracy of 97.08%, and an area under the curve (AUC) of 99.23%, which is a significant improvement over the unimodal baselines. Future research will focus on enhancing the proposed model by investigating transformer-based architectures to achieve deeper and more contextualized representation learning.