Reducing Conversational Agents’ Overconfidence Through Linguistic Calibration

Abstract While improving neural dialogue agents’ factual accuracy is the object of much research, another important aspect communication, less studied in setting dialogue, transparency about ignorance. In this work, we analyze to what extent state-of-the-art chit-chat models are linguistically calibrated sense that their verbalized expression doubt (or confidence) matches likelihood model’s responses factually incorrect correct). We find these poorly calibrated, yet show correctness can accurately be predicted. By incorporating such metacognitive features into training a controllable generation model, obtain agent with greatly improved linguistic calibration.