[2F4-OS-33] OS-33

2:30 PM - 2:45 PM JST(5:30 AM - 5:45 AM UTC)

[2F4-OS-33-05] Predicting Contact-Exchange Intent in Speed-Dating Dialogues via LLM-Based Important-Segment Selection

〇Yuriko Kikuchi¹, Takato Hayashi¹, Ryusei Kimura¹, Ryo Ishii², Shogo Okada¹ (1. Japan Advanced Institute of Science and Technology, 2. NTT, Inc)

Keywords:

Social Signal Processing,Multimodal Learning,Explainable AI

Speed-dating dialogues are weakly supervised with only a dialogue-level label, making it hard to identify decision-relevant segments. We propose an LLM-based selector that chooses one important one-minute segment per dialogue and predicts contact-exchange intent (match/unmatch) from multimodal features of that segment. We trained a Random Forest on features from all ten one-minute segments and, at test time, used only the LLM-selected segment as input. With 10-fold cross-validation, our method achieved macro-F1 comparable to the best fixed-window baseline while providing dialogue-specific important segments. We further validated the selection via comparisons with random selection, segment removal, and before/after segment evaluations. J-LIWC analysis suggests the LLM prefers segments richer in drives and auxiliary verbs, and poorer in third-person pronouns and fillers.