BnClinical-Sum: benchmarking datasets for Bangla long & short clinical dialogue summarization

Abstract

Despite significant improvements in the general-purpose text summarization task in the past decade, clinical conversion summarization is going through a tough time due to a lack of initiative to provide open-source datasets to the NLP community. In this work, we are presenting the first long and short Bangla Clinical Dialogue to Note Summarization datasets: BnClinical-Sum. Long conversations are detailed conversations with additional medical history. For the long dialogue dataset, we have accumulated around 207 pairs of full conversations and notes. Each note consists of in-depth discussions on previous medical histories, family medical records, and a wide variety of other topics. For the short dialogue version, our dataset consists of 1701 real-life short manually translated clinical conversations and their corresponding notes. The short dialogue dataset consists of subsets of long dialogue where each dialogue snippet addresses one sub-topic like previous medical histories, family medical records, etc. Those conversations are from 20 different categories like labs, assessments, plans, etc. Owing to demonstrating the efficacy of both datasets, we have trained our datasets on current state-of-the-art text summarization and text-to-text generative models to provide a solid benchmark for clinical conversion summarization tasks.