AffectLLM: A Large Language Model for Rating the Affective Quality of Representations

Affective computing presents significant challenges, particularly in the context of human-machine interactions. Traditional approaches often involve developing data-driven models that rely on physiological and image-based data to infer the emotional states of individuals during interactions with machines. However, these models frequently lack contextual relevance and perform poorly in complex scenarios involving conflicting cues. Additionally, they suffer from weak generalization due to the absence of structured methodologies that disentangle interconnected external factors. To address these limitations, we propose leveraging the Affective Quality of Representations (AFF) score, a standardized metric for assessing human affective responses across varied contexts. The AFF score is derived from expert ratings of participant narratives elicited by questions or visual cards, following established SCORS-G guidelines. The primary challenge of this approach is its reliance on trained psychology experts for scoring. To mitigate this, we developed the AffectLLM, a large language model designed to predict AFF scores efficiently and support the analysis of human-machine interactions. A strong contribution of AffectLLM is providing a structured metric that can guide the design of humanmachine interaction environments that promote more positive affect quality outcomes.