Wikitology: a Novel Hybrid Knowledge Base Derived from Wikipedia

Title of dissertation: WIKITOLOGY: A NOVEL HYBRID KNOWLEDGE BASE DERIVED FROM WIKIPEDIA Zareen Saba Syed, Doctor of Philosophy, 2010 Dissertation directed by: Professor Timothy W. Finin Department of Computer Science and Electrical Engineering World knowledge may be available in different forms such as relational databases, triple stores, link graphs, meta-data and free text. Human minds are capable of understanding and reasoning over knowledge represented in different ways and are influenced by different social, contextual and environmental factors. By following a similar model, we have integrated a variety of knowledge sources in a novel way to produce a single hybrid knowledge base i.e., Wikitology, enabling applications to better access and exploit knowledge hidden in different forms. Wikipedia proves to be an invaluable resource for generating a hybrid knowledge base due to the availability and interlinking of structured, semi-structured and un-structured encyclopedic information. However, Wikipedia is designed in a way that facilitates human understanding and contribution by providing interlinking of articles and categories for better browsing and search of information, making the content easily understandable to humans but requiring intelligent approaches for being exploited by applications directly. Research projects like Cyc [61] have resulted in the development of a complex broad coverage knowledge base however, relatively few applications have been built that really exploit it. In contrast, the design and development of Wikitology KB has been incremental and has been driven and guided by a variety of applications and approaches that exploit the knowledge available in Wikipedia in different ways. This evolution has resulted in the development of a hybrid knowledge base that not only incorporates and integrates a variety of knowledge resources but also a variety of data structures, and exposes the knowledge hidden in different forms to applications through a single integrated query interface. We demonstrate the value of the derived knowledge base by developing problem specific intelligent approaches that exploit Wikitology for a diverse set of use cases, namely, document concept prediction, cross document co-reference resolution defined as a task in Automatic Content Extraction (ACE) [1], Entity Linking to KB entities defined as a part of Text Analysis Conference Knowledge Base Population Track 2009 [65] and interpreting tables [94]. These use cases directly serve to evaluate the utility of the knowledge base for different applications and also demonstrate how the knowledge base could be exploited in different ways. Based on our work we have also developed a Wikitology API that applications can use to exploit this unique hybrid knowledge resource for solving real world problems. The different use cases that exploit Wikitology for solving real world problems also contribute to enriching the knowledge base automatically. The document concept prediction approach can predict inter-article and category-links for new Wikipedia articles. Cross document co-reference resolution and entity linking provide a way for specifically linking entity mentions in Wikipedia articles or external articles to the entity articles in Wikipedia and also help in suggesting redirects. In addition to that we have also developed specific approaches aimed at automatically enriching the Wikitology KB by unsupervised discovery of ontology elements using the inter-article links, generating disambiguation trees for entities and estimating the page rank of Wikipedia concepts to serve as a measure of popularity. The set of approaches combined together can contribute to a number of steps in a broader unified framework for automatically adding new concepts to the Wikitology knowledge base. WIKITOLOGY: A NOVEL HYBRID KNOWLEDGE BASE DERIVED FROM WIKIPEDIA