A robust approach to extract biomedical events from literature

Abstract

Motivation: The abundance of biomedical literature has attracted significant interest in novel methods to automatically extract biomedical relations from the literature. Until recently, most research was focused on extracting binary relations such as protein–protein interactions and drug–disease relations. However, these binary relations cannot fully represent the original biomedical data. Therefore, there is a need for methods that can extract fine-grained and complex relations known as biomedical events. Results: In this article we propose a novel method to extract biomedical events from text. Our method consists of two phases. In the first phase, training data are mapped into structured representations. Based on that, templates are used to extract rules automatically. In the second phase, extraction methods are developed to process the obtained rules. When evaluated against the Genia event extraction abstract and full-text test datasets (Task 1), we obtain results with F-scores of 52.34 and 53.34, respectively, which are comparable to the state-of-the-art systems. Furthermore, our system achieves superior performance in terms of computational efficiency.