How You Label Strategy Without Labels: Oracle Inpainting and Counterfactuals

Explain the dataset problem and the oracle approach: analyze full dialogues and outcomes, infer which primitives were effective, then generate grounded reasons and scores. Introduce counterfactual scenario augmentation.

The fundamental dataset problem in persuasion modeling is that real dialogues are not labeled with strategy effectiveness. We have the transcripts; we rarely know which specific moves drove which outcomes, or what would have happened if a different strategy had been used at a given turn.

Oracle inpainting addresses this by working backward from outcomes: given a successful persuasion event, infer which primitive combinations at which turns were causally relevant, then generate grounded labels and effectiveness scores for those turns. Counterfactual augmentation extends this by generating alternative scenarios where different primitives were used at the same turns, teaching the model conditional causality rather than just correlation.