Bio-Molecular Event Extraction Using Classifier Ensemble-of-Ensemble Technique

Abstract

Biomedical literature is a vast repository of critical research information that is highly desirable to be understood and made use of in a timely manner by scientists in biology and life sciences. Every day greater than 3000 articles are being published in biomedical journals. PubMed, the largest database for biomedical research, presently has more than 29.1 million records, growing exponentially at more than 1.2 million records annually. Advanced Natural Language Processing (NLP) methods can contribute significantly to the processing of this huge volume of literature. This research proposes a novel way of finding expressions relevant to bio-molecular events, interpret them, categorize them into predefined categories, and eventually add claims so as to be understood. The proposed methodology uses supervised machine learning techniques such as well-known classification algorithms like Support Vector Machine, Decision tree, K-nearest neighbor, and their kernel variants as base-learners in the first-layer ensemble. We then combine the output of these base-learners in a second layer, the meta-learner, or ensemble-of-ensemble as a novelty to enhance the overall event extraction accuracy. MATLAB-based simulation and experiments on benchmark dataset of BioNLP-2013 shared task results in an overall F-score of 66.34%. Comparing with existing systems, namely, TEES 2.1, EVEX and BioSEM ranked #1 at the BioNLP Shared task 2013 challenge with an F-score of 51%, FAUST and EventMine with 56.04% and 57.28%, respectively, at BioNLP Shared task 2011 challenge indicate that the proposed system attains an improvement of >10% along with state-of-the-art performance.