Novel Approaches for Relation Extraction in Biomedical Domain

Relation extraction an important field in Biomedical Natural Language Processing is the study of identifying relations between entity mentions. The extraction of relation instances over multiple sentence mention levels (intra- and inter-sentence levels) has been a challenge. In the intra-sentence level, the mention of a pair of entity is found in a single sentence, whereas in the inter-sentence level, they are found in spanning neighbouring sentences. The variations in the level of extractable information and performance from these levels have been a reason for this challenge. In this thesis, we tackled this challenge by carefully examining the stages of text processing and relation instance construction of the candidate relation instances across the multiple sentence levels and further performed a combination of the relation instances over these mention levels in order improve the performance of the system. In the text processing stage, we performed sentence simplification after the sentences have been segmented in order to improve the information extracted through a dependency parse tree. During the extraction of the candidate relation instances, we applied some sentence structures and rules to help improve the level of the types of candidates selected. We performed relation extraction using two systems. We developed a system that employs an optimization technique namely genetic algorithm, to combine the output of the classifiers trained using the candidate relation instances from both levels. We introduce the novel approach of using two decision-making under uncertainty techniques for our classifier selection. The other system is based on an ensemble of two machine learning algorithms. We performed relation extraction by employing the candidate relation instances from the two levels in two forms. Firstly, the instances are merged after they have been classified individually, and secondly, the instances are merged before the classification. The system then introduces the novel use of a maximum probability-based voting algorithm to combine the results generated from these two forms. All the experiments in this study are performed using the BioCreative V chemical disease relation dataset which is the most comprehensive dataset in the domain. Keywords: Classifier Ensemble, Decision-Making Techniques, Genetic Algorithms, Optimization Techniques, Relation Extraction, Text Mining.