Using conditional random fields for result identification in biomedical abstracts

The abstracts of biomedical papers usually contain three sections: objective, methods, and results-conclusion. The results-conclusion section is the most important because it usually describes the main contribution of a paper. Unfortunately, not all biomedical journals follow this three-section format. In this paper, we propose a machine learning (ML) based approach to automatically identify the results-conclusion section. The results-conclusion section identification problem is formulated as a sequence labeling task. Four feature sets, including Position, Named Entity, Tense, and Word Frequency, are employed with Conditional Random Fields (CRFs) as the underlying ML model. The experiment results show that the proposed approach can achieve F-measure, precision, and recall of 97.08%, 96.63% and 97.53%, respectively.