Biomolecular Event Extraction using a Stacked Generalization based Classifier

In this paper we propose a stacked generalization (or stacking) model for event extraction in bio-medical text. Event extraction deals with the process of extracting detailed biological phenomenon, which is more challenging compared to the traditional binary relation extraction such as protein-protein interaction. The overall process consists of mainly three steps: event trigger detection, argument extraction by edge detection and finding correct combination of arguments. In stacking, we use Linear Support Vector Classification (Linear SVC), Logistic Regression (LR) and Stochastic Gradient Descent (SGD) as base-level learning algorithms. As meta-level learner we use Linear SVC. In edge detection step, we find out the arguments of triggers detected in trigger detection step using a SVM classifier. To find correct combination of arguments, we use rules generated by studying the properties of bio-molecular event expressions, and form an event expression consisting of event trigger, its class and arguments. The output of trigger detection is fed to edge detection for argument extraction. Experiments on benchmark datasets of BioNLP2011 show the recall, precision and Fscore of 48.96%, 66.46% and 56.38%, respectively. Comparisons with the existing systems show that our proposed model attains state-of-the-art performance.