Expectation Maximization Algorithm for Domain Specific Ontology Extraction

Learning ontology from unstructured text is a challenging task. Over the years, a lot of research has been done to predict ontological relation between a pair of concepts. However all these measures predict relation with a varying degree of accuracy. There has also been work on learning ontology by combining evidences from heterogeneous sources, but most of these algorithms are ad hoc in nature. In this paper we investigate wide range of evidences to predict relation between a pair of concepts and propose a standardized Expectation Maximization algorithm to construct domain specific ontology. The proposed approach is completely unsupervised and does not require any seed terms or human intervention. In addition, the proposed approach can also be easily adopted for any language. We have conducted our experiments for two languages Hindi and English and for two domains Health and Tourism. The average F-Score observed in all experiments is above 0.60.