Maximum Entropy Learning Model for Biomedical Mention Classification

I, Ekaterina Shutova, of Hughes Hall, being a candidate for the M.Phil in Computer Speech, Text and Internet Technology, hereby declare that this dissertation and the work described in it are my own work, unaided except as may be specified below, and that the dissertation does not contain material that has already been used to any substantial extent for a comparable purpose. Abstract " Genes can be named on the basis of a mutant phenotype or on the basis of the predicted protein product or RNA product " (Genetic Nomenclature, J. Hodgkin). This results in the fact that in the biomedical literature same gene names tend to be used to refer to more than one entity. It is very common for a particular gene name to refer to the protein produced, or some other product. Therefore, when performing information extraction tasks identifying gene names is not sufficient and it is necessary to distinguish between all biological entities for which the same gene name is used, which is the key objective of this project. The proposed approach consists of training a Maximum Entropy classifier for the task in question, using a large automatically created biomedical corpus, as opposed to other systems being trained on expensive manually annotated data. First, a rule-based recognizer is developed using information from ontology that classifies biological mentions into seven biotypes. Subsequently this baseline system is applied to a large corpus of biomedical text in order to tag it. The resulting corpus constitutes the training data for the machine learning system. As expected, the bootstrapped maximum entropy classifier trained on automatically created data performs better than the baseline recognizer itself, which proves the efficiency of the adopted approach. Subsequently, machine learning techniques investigated and developed within the framework of the project can be applied to text mining in any domain.