Modality Discrepancy Transformer for Recognizing Affective Ambivalence

The Modality Discrepancy Transformer (MDT) identifies ambivalence and hesitancy by detecting contradictions across facial, vocal, and linguistic signals. It utilizes a 9-token representation including absolute-difference and Hadamard-product features to capture cross-modal disagreement through self-attention and LoRA fine-tuning.