Enhancing Biomedical Named Entity Classification Using Terabyte Unlabeled Data

摘要

This paper presents a semi-supervised learning method to enhance biomedical named entity classification using features generated from labeled and terabyte unlabeled data, called Feature Coupling Degree (FCD) features. Highly discriminative context words are obtained from labeled free text using Chi-square method and queries formed by combining the named entity and context words are retrieved by search engine. Then the retrieved web page counts are converted into binary features by discretization. We investigate the effect of this type of feature in a biomedical corpus generated from several online resources. Support Vector Machine (SVM) is used as classifier and the performances of different features with various kernels and discretization methods are compared. The results show that the method enhances the classification performance especially for Out-of-Vocabulary (OOV) terms and relative small size of training data. In addition, only using FCD features with polynomial kernels, the performance is competitive to classical features.