Reducing conversational agents' overconfidence through linguistic calibration
This work addresses the problem of conversational agents' overconfidence, which is crucial for improving user trust and interaction quality for anyone interacting with these systems. It's an incremental improvement in a specific aspect of dialogue generation.
This paper investigates the linguistic calibration of state-of-the-art chit-chat models, finding them to be poorly calibrated in expressing doubt or confidence relative to factual correctness. By integrating metacognitive features into a controllable generation model, the authors achieve a dialogue agent with greatly improved linguistic calibration.
While improving neural dialogue agents' factual accuracy is the object of much research, another important aspect of communication, less studied in the setting of neural dialogue, is transparency about ignorance. In this work, we analyze to what extent state-of-the-art chit-chat models are linguistically calibrated in the sense that their verbalized expression of doubt (or confidence) matches the likelihood that the model's responses are factually incorrect (or correct). We find that these models are poorly calibrated, yet we show that likelihood of correctness can accurately be predicted. By incorporating such metacognitive features into the training of a controllable generation model, we obtain a dialogue agent with greatly improved linguistic calibration. While improving neural dialogue agents' factual accuracy is the object of much research, another important aspect of communication, less studied in the setting of neural dialogue, is transparency about ignorance. In this work, we analyze to what extent state-of-the-art chit-chat models are linguistically calibrated in the sense that their verbalized expression of doubt (or confidence) matches the likelihood that the model's responses are factually incorrect (or correct). We find that these models are poorly calibrated, yet we show that likelihood of correctness can accurately be predicted. By incorporating such metacognitive features into the training of a controllable generation model, we obtain a dialogue agent with greatly improved linguistic calibration.