Ambivalence and Hesitancy Recognition

Ambivalence and hesitancy recognition studies how to identify mixed feelings or uncertainty, often from short video interactions relevant to counseling and digital health. Approaches combine transcript and language models with speech prosody, facial expressions or action units, and visual context, using cross-attention, late fusion, or features that explicitly measure disagreement between modalities. The BAH dataset and ABAW challenge provide common video-level evaluation settings, where text can be a strong signal and multimodal fusion may add complementary evidence. A central challenge is making results reliable across speakers and datasets: limited labels, subtle behaviors, and calibration choices can make apparent gains hard to generalize.