Leveraging Machine-generated Rationales for Conversational Forecasting
Track: Preliminary Work Track
Keywords: conversational forecasting, rationales
TL;DR: We use machine generated rationales to improve conversational forecasting success
Abstract: Predicting outcomes in multi-turn dialogues is challenging due to the implicit nature of decision-making and the evolving dynamics between participants. In this work, we explore whether LLM-generated rationales can enhance the accuracy and generalizability of outcome prediction in task-oriented dialogues. We evaluate zero-shot in-context learning models on the Craigslist Bargain dataset, testing their ability to predict final sale prices at different dialogue checkpoints in absence and presence of rationales. Preliminary results with metrics such as RMSE and Pearson correlation suggest that rationale-augmented models better capture negotiation strategies and concession patterns, improving early-stage prediction accuracy.
Submission Number: 18
Loading