Recognizing Biomedical Named Entities in the Absence of Human Annotated Corpora

Biomedical named entity recognition is an important task in biomedical text mining. Currently the dominant approach is supervised learning, which requires a sufficiently large human annotated corpus for training. In this paper, we propose a novel approach aimed at minimizing the annotation requirement. The idea is to use a dictionary which is essentially a list of entity names compiled by domain experts and sometimes more readily available than domain experts themselves. Given an unlabelled training corpus, we label the sentences by a simple dictionary lookup, which provides us with highly reliable but incomplete positive data. We then run a SVM-based self-training process in the spirit of semi-supervised learning to iteratively learn from the positive and unlabelled data to build a reliable classifier. Our evaluation on the BioNLP-2004 shared task data sets suggests that the proposed method can be a feasible alternative to traditional approaches when human annotation is not available.