Training Set Expansion Using Word Embeddings for Korean Medical Information Extraction

Abstract

Entity recognition is an essential part of a task-oriented dialogue system and is considered as a sequence labeling task. However, constructing a training set in a new domain is extremely expensive and time-consuming. In this work, we propose a simple framework to exploit neural word embeddings in a semi-supervised manner to annotate medical named entities in Korean. The target domain is the automatic medical diagnosis, where disease name, symptom, and body part are defined as the entity types. Different aspects of the word embeddings such as embedding dimension, window size, models are examined to investigate their effects on the final performance. An online medical QA data has been used for the experiments. With a limit number of pre-annotated words, our framework could successfully expand the training set.