WikiMatch results for OEAI 2012

WikiMatch is a matching tool which makes use of Wikipedia as an external knowledge resource. The overall idea is to search Wikipedia for a given concept and retrieve all pages describing the term. If there is a large amount of common pages for two terms, then the concepts will have similar semantics. We make also use of the inter-language links between Wikipedias in different languages to match multilingual ontologies. The results show that this simple idea can keep up with state of the art tools. Moreover, the results on the Multifarm track depend on the Wikipedia’s number of articles as well as the link amount to the Wikipedia of the other natural language to match. The growth of Wikipedia will thus help this matcher to improve the matching quality. 1 Presentation of the system 1.1 State, purpose, general statement WikiMatch is an element-level ontology matching tool. It uses Wikipedia as a huge background knowledge to find out, how similar two concepts are. The algorithm extracts all labels, comments, and URI fragments, and uses Wikipedia’s search function to retrieve an set of articles related to that term. If the intersection between such two sets is high, then we assume that the terms have something in common and are related to each other. To also deal with multilingual ontologies, all language links of the returned articles are requested as a second step. For each language, the Jaccard coefficient of the two sets of articles retrieved is computed, as equation (1) shows. sim(t1, t2) := maxti∈{label(ci),fragment(ci),comment(ci)},i∈{1,2} #(S(t1) ∩ S(t2)) #(S(t1) ∪ S(t2)) (1) For the terms t1 and t2 the resulting article set S is computed. The maximum over all labels, comments and URI fragments are then the similarity measure for these terms. If Wikipedia returns an suggestion for the term, a new query is made with this new search term. This is typically the case when entering a misspelled term in the search. An overview of the entire system is shown in Fig. 1. 1.2 Specific techniques used Our first test was to search for the whole term in Wikipedia. We call this approach simple search. As a result the precision is high in contrast to the recall which is very low. To