MedExpDial: Machine-to-Machine Generation of Explanatory Dialogues for Medical QA

We describe a pilot study on generating synthetic explanatory dialogues for the medical domain, based on a pre-existing medical dataset of multiplechoice questions with human-written explanations. We use an instruction-tuned large language model (LLM) to generate dialogues between a medical student and a teacher/doctor helping answer questions about clinical cases. We inject varying degrees of background knowledge into the teacher prompt and analyze the effectiveness of these dialogues in terms of whether the student is able to get to the correct answer and in how many turns. This method has potential applications in developing and evaluating argument-based explanations.