Textualized and Feature-based Models for Compound Multimodal Emotion Recognition in the Wild
The study compares two methods for recognizing compound emotions in videos: feature-based and textualized models. Textualization augments data with emotional cues to help large language models understand connections between video, audio, and text. Results show textualization performs better when video transcripts are rich, suggesting it is most effective when high-quality text descriptions are available.