Enhancing Medical Named Entity Recognition with Features Derived from Unsupervised Methods

Abstract

A study of the usefulness of features extracted from unsupervised methods is proposed. The usefulness of these features will be studied on the task of performing named entity recognition within one clinical sub-domain as well as on the task of adapting a named entity recognition model to a new clinical sub-domain. Four named entity types, all very relevant for clinical information extraction, will be studied: Disorder, Finding, Pharmaceutical Drug and Body Structure. The named entity recognition will be performed using conditional random fields. As unsupervised features, a clustering of the semantic representation of words obtained from a random indexing word space will be used.