Evaluating VAD Similarity for Emotion-Cause Retrieval

Abstract

Abstract We add VAD-based similarity alongside semantic similarity in a custom re-ranker to explore the impact of emotional similarity on an emotion-cause retrieval task for a curated subset of the RECCON dataset (99 conversations). Recall@1, Recall@3, and MRR were tracked across five unique weight distributions. Our empirical results suggest that the semantic-only baseline achieved the best Recall@1 and MRR scores, whereas the 0.75/0.25 semantic-VAD weight split achieved the best Recall@3 score. The VAD-only distribution, however, had the lowest Recall@1, Recall@3, and MRR across all distributions. We further observe that VAD-based retrieval may be better suited to cases of emotional continuity than cases of emotional transition, suggesting that emotional similarity alone might not capture emotion-cause relevance.