Maximum Entropy Learning Model for Biomedical Semantic Type Induction

In the biomedical literature the gene names tend to be used to refer to biomedical entities other than genes. Therefore, when performing information extraction tasks it is necessary to distinguish between all such biomedical entities. Our approach consists of training a Maximum Entropy classifier for this task using a large automatically-created training corpus, as opposed to manually annotated data. In order to create the training corpus, a rulebased baseline tagger is developed using the information from the Sequence Ontology that classifies biomedical entities into seven biotypes. Subsequently, this tagger is applied to a large corpus of biomedical text in order to annotate it, thus generating the training data for the machine learning system. The results obtained show that the maximum entropy classifier trained on automatically created data performs better than the baseline tagger itself, which demonstrates the efficiency of the adopted approach. The developed machine learning techniques are domain-independent and can be applied to text mining in any field.