Dual-Branch Feature Extraction via Discrepancy-Aware Fusion with Evidential Deep Learning for Sarcasm Detection
Takato Ueno, Keito Inoshita · 2025
Sarcasm detection remains a challenging task for Large Language Models (LLM) due to the difficulty in capturing the semantic gap between the literal meaning of an utterance and the speaker’s intended sentiment. Although traditional deep learning approaches have been employed to address this issue, they often struggle to handle the ambiguity and subtlety inherent in sarcastic expressions, making it difficult to properly distinguish between literal content and underlying emotions. Moreover, since sarcasm heavily depends on context and the subjective interpretation of the listener, models that cannot assess the confidence of their predictions are more likely to make incorrect judgments in ambiguous cases. Therefore, there is a growing need for models that not only capture the semantic discrepancy but also evaluate the reliability of their predictions. In this study, we propose Dual-Branch feature extraction via Discrepancy-Aware fusion with Evidential Deep Learning (DBDA-EDL), a deep learning-based approach that explicitly captures the semantic discrepancy characteristic of sarcasm and utilizes evidence-based learning. The proposed model consists of three key components: Dual-Branch Feature Extraction (DBFE), which independently extracts literal meaning and emotional expressions; Discrepancy-Aware Fusion (DAF), which emphasizes branch-wise differences using a mismatch computation mechanism and a bilinear gate to model semantic discrepancy; and Evidential Deep Learning (EDL), which assigns low confidence to ambiguous samples, thereby improving estimation under uncertainty. Experimental results demonstrate that DBDA-EDL outperforms other baseline models, achieving superior classification performance and learning efficiency compared to previous models. These findings validate DBDA-EDL as a highly accurate and reliable approach for sarcasm detection.