Bootstrapping and Evaluating Named Entity Recognition in the Biomedical Domain

We demonstrate that bootstrapping a gene name recognizer for FlyBase curation from automatically annotated noisy text is more effective than fully supervised training of the recognizer on more general manually annotated biomedical text. We present a new test set for this task based on an annotation scheme which distinguishes gene names from gene mentions, enabling a more consistent annotation. Evaluating our recognizer using this test set indicates that performance on unseen genes is its main weakness. We evaluate extensions to the technique used to generate training data designed to ameliorate this problem.