Markov Logic for Ontology based Information Extraction

Abstract

Abstract Today’s world, Internet has become a fast and efficient information provider although the relevancy or accuracy of the information found is not guaranteed. Web itself presents numerous problems mainly due to its heterogeneous nature with respect to semantics and representation formalisms, high dynamicity and the overwhelming size of the resources available. Therefore, encoding semantics of web documents in a formal way is a necessity in effective information gathering. Ontology plays a vital role in enhancing the semantics of natural language documents in machine-readable form on the semantic web. In ontology construction and in linking terms in web documents with appropriate concepts in Ontologies, terms and relationships between them should be extracted. Successful information extraction, for ontology construction needs to focus on natural language sentences for identifying the concepts the document entities represented and their relationships. As a result of the efforts made by semantic web researchers, numerous established techniques and tools are available mainly for the extraction of entities which form basic constituents of ontology, the concepts. However, the associated Relation extraction is yet to be addressed extensively. Therefore, the present work concentrates on extracting domain specific entities and generating relation-extraction-rules to extract relations for ontological structures. The presented method exploits the existing techniques for entity extraction and introduces a novel approach for relation extraction based on a set of rules specifying the dependencies of entities in natural language sentences. Adaptation of Inductive Logic Programming to generate relation-extraction-rules from language dependency clauses and modeling them on Markov Logic Network environment for statistical relation extraction by using a domain independent approach, distinguishes the present work from previous work in the area of information extraction. The evaluation of the system shows the effectiveness of the proposed method in domain specific information extraction and in document classification as a proof of concept. Document classification shows a high accuracy in all classifications especially with 100% precision on number of occasions in the selected domain. Furthermore, this research stresses the importance of evolving training corpus automatically in order to minimize the manual involvement in supervised learning in creating a large amount of training data.