Leveraging rich annotations to improve learning of medical concepts from clinical free text

Information extraction from clinical free text is one of the key elements in medical informatics research. In this paper we propose a general framework to improve learning-based information extraction systems with the help of rich annotations (i.e., annotators provide the medical assertion as well as evidences that support the assertion). A special graphical interface was developed to facilitate the annotation process, and we show how to implement this framework with a state-of-the-art context-based question answering system. Empirical studies demonstrate that with about 10% longer annotation time, we can significantly improve the accuracy of the system. An approach to provide supporting evidence for test documents is also briefly discussed with promising preliminary results.