Evolutionary Approach for Classifier Ensemble: An Application to Bio-molecular Event Extraction

Abstract

The main goal of Biomedical Natural Language Processing (BioNLP) is to capture biomedical phenomena from textual data by extracting relevant entities, information and relations between biomedical entities (i.e. proteins and genes). Most of the previous works focused on extracting binary relations among proteins. In recent years, the focus is shifted towards extracting more complex relations in the form of bio-molecular events that may include several entities or other relations. In this paper we propose a classifier ensemble based on an evolutionary approach, namely differential evolution that enables extraction, i.e. identification and classification of relatively complex bio-molecular events. The ensemble is built on the base classifiers, namely Support Vector Machine, nave-Bayes and IBk. Based on these individual classifiers, we generate 15 models by considering various subsets of features. We identify and implement a rich set of statistical and linguistic features that represent various morphological, syntactic and contextual information of the candidate bio-molecular trigger words. Evaluation on the BioNLP 2009 shared task datasets show the overall recall, precision and F-measure values of 42.76%, 49.21% and 45.76%, respectively for the three-fold cross validation. This is better than the best performing SVM based individual classifier by 4.10 F-measure points.