Coreference resolution on entities and events for hospital discharge summaries

Abstract

The wealth of medical information contained in electronic medical records (EMRs) and Natural Language Processing (NLP) technologies that can automatically extract information from them have opened the doors to automatic patient-care quality monitoring and medical- assist question answering systems. This thesis studies coreference resolution, an information extraction (IE) subtask that links together specific mentions to each entity. Coreference resolution enables us to find changes in the state of entities and makes it possible to answer questions regarding the information thus obtained. We perform coreference resolution on a specific type of EMR, the hospital discharge summary. We treat coreference resolution as a binary classification problem. Our approach yields insights into the critical features for coreference resolution for entities that fall into five medical semantic categories that commonly appear in discharge summaries.