Data-Driven Decoding of Russell's Circumplex Model of Affect

This paper presents a novel framework for validating emotional representations using deep learning, specifically examining how Transformer embeddings, including RoBERTa for text and wav2vec 2.0 for audio, align with Russell's circumplex model of affect. The study demonstrates that a multimodal fusion architecture achieves perfect topological alignment with human emotion mappings, indicating that the latent spaces of these models intrinsically encode emotional structures. This finding is significant for practitioners as it provides a data-driven approach to understanding and improving emotion recognition systems in AI, enhancing their interpretability and relevance in affective computing applications.