Home → News and Events → Events Calendar → Seminar
Question Answering from Electronic Health Records
When: 11:00AM - 12:00PM , October 17, 2019
Speaker: Bhanu Pratap Singh
Past decade has seen the development and release of several questions answering datasets (such as SQuAD, HotpotQA, emrQA, QuAC, Natural questions), models (Mem Nets, BiDAF, BERT, ERNIE) and their respective leaderboards. These models at times are able to achieve near-human performance but that is not the case for domain heavy question answering models. In medical domain, another hurdle is to work with large context and noisy natural language paragraphs from medical records. In our recent work, we propose a multi-task learning model that works at sentence level while also learning the evidence sentences from the whole medical records. We show the efficacy of our model for an important task in the biomedical domain to obtain the causal inference between a medication and its adverse drug reactions (ADRs). I would also briefly discuss my recent work at IBM research, where we worked on integrating question logical forms in training QA models on unstructured clinical notes.
Bhanu is a PhD student at UMass Amherst and works with Prof. Hong Yu in the BioNLP lab. His research focuses on building machine learning methods, especially in the domain of natural language processing (NLP), to advance the knowledge in healthcare. Bhanu is currently focussing on question answering from electronic health records (EHRs) and his recent work was published at SIGKDD’19 and MLHC’19. Before his grad studies, he worked as a Research Engineer for 3 years in India where his work was focused on the intersection of human computer interface (HCI) and machine learning.