Cross-Cultural Generalization of Interpersonal Influence Models in Dyadic Conversation

Abstract

The ACII 2026 DaiKon Influence sub-challenge evaluates segment-level prediction of affective interpersonal influence in dyadic conversation. We study whether Influence models trained on one country-defined cultural group generalize to others, restricting analysis to five countries in the current test-set result bundle: Germany, the United States, Spain, the Netherlands, and Poland. The task is formulated as multi-output regression over ten dimensions: anger, anxiety, uncertainty, confusion, doubt, boredom, surprise, curiosity, joy, and amusement. Using target-speaker country metadata, we train one source model per country and evaluate it on every target country. With a simple MLP over mean-pooled Whisper and FaceNet segment features, mean within-country CCC is 0.272 versus 0.128 for off-diagonal cross-country transfer. Pooled multicultural training reaches mean CCC 0.210, while few-shot target-country adaptation improves mean CCC from 0.180 with no target labels to 0.273 with 25% of target-country training rows. Results indicate that Influence prediction contains transferable structure but also substantial country- or domain-specific effects.