Researchers have developed an audio-text system for the 11th ABAW Competition's Ambivalence/Hesitancy Video Recognition Challenge. This system, which omits visual data, processes videos in 5-second windows, combining prosodic audio features, RoBERTa embeddings, and psycholinguistic indicators of uncertainty. The audio and text components are fused using temporal cross-attention, with support features integrated before pooling to weigh window importance. An ensemble of five independently trained models achieved an average precision of 0.875 and a macro-F1 score of 0.72 on the development set. AI
IMPACT This system advances techniques for recognizing nuanced human emotions like ambivalence and hesitancy using multimodal AI, potentially improving human-computer interaction and sentiment analysis tools.
RANK_REASON The cluster describes a research paper detailing a new system for a specific recognition challenge, with code made publicly available.
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
