Evaluating Previous-Partner Context for Affective Influence Prediction in Dyadic Conversation

Abstract

Automatic understanding of affect in conversation requires models to represent both a speaker’s behaviour and the surrounding interaction. We study whether the most recent partner segment preceding the current target segment adds predictive value in the DaiKon Influence task. Using mean-pooled Whisper-small representations and a two-hidden-layer multilayer perceptron, we compare five controlled input conditions across seeds 7, 42, and 123. On validation, the true-partner condition T4 obtains 0.428 ± 0.002 mean concordance correlation coefficient (CCC), compared with 0.405 ± 0.003 for current target audio (T1), 0.422±0.003 for target history (T3), and 0.411±0.001 for a random-partner control (T5). The T4>T3 validation ordering holds for all three seeds. On the held-out test split, T4 obtains 0.420±0.001, compared with 0.407±0.004 for T1, 0.420±0.002 for T3, and 0.403 ± 0.001 for T5. Thus, T4 and T3 tie to three decimal places on test, while T4 obtains higher test CCC than the current-only and random-partner controls. The evidence supports a small, seed-stable validation gain beyond target history and a literal held-out test advantage over T5, but does not establish causal interpersonal influence.