Question Answering System for Healthcare Information based on BERT and GPT

Recently, demand for telemedicine counseling has been on the rise. Although telemedicine is legally prohibited in Korea, the number of healthcare-related questions on online question and answer platforms is steadily increasing. However, doctors have to manually write answers to similar questions, which is time-consuming and inefficient. This study implemented a healthcare information provision question and answer system using an algorithm that combines BERT and GPT-2 structures. Intellectual Q&A pairs were used as training data, and the BERT-GPT algorithm experimented with creating sentences that provide healthcare information such as related medical subjects and suspected diseases as answers to healthcare-related questions. As a result of BERT-GPT2 algorithm gets worse PPL score and loss than the baseline model, but better score in the qualitative assessment. This result means BERT-GPT2 understands the conversation intention better.