Developing ML-based Systems to Extract Medical Information from Japanese Medical History Summaries

With the increase of the number of medical records written in an electronic format, natural language processing techniques in the medical domain have become more and more important. For the purpose of the development and evaluation of machine learning-based systems to extract medical information, we recently participated in the NTCIR-10 MedNLP task. The task focused on Japanese medical records and aimed at evaluating different information extraction techniques on the common data set provided by the organizers. We implemented our baseline system based on structured perceptron and have developed its extensions. In this paper, we describe our systems and report on the evaluation of and the analysis on their performance.