Question Answering Chatbots for Biomedical Research using Transformers

Professionals as well as the general public need effective help to access, understand and consume complex biomedical concepts. The existence of an interaction environment capable of automatically processing such information - thus replacing human intervention - such as chatbots, is however challenging. In this paper we propose a method of utilizing chatbots in the domain of biomedicine. In the implementation we choose to incorporate the BERT algorithm, so as to adopt a modern technique for natural language processing tasks. We use several pre-trained models (RoBERTa, XLM-R, BERT Large, and BioBert) in order to evaluate their ability to back the chatbot infrastructure. The data is retrieved from the PubMed repository, with the final set being formed into full sentences or potential chatbot responses, thus preserving their conceptual meaning. Response selection is performed using similarity metrics and F-score. The results create a ranking of the models placing related ones closely, recognizing the ability to always answer each question and highlighting the importance of the training previously applied to them. These are compared to the Count Vectorizer technique, which appears to perform better, but with several weaknesses, as many questions could not be answered.