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.
1.1
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. 1.2
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 O1 read ontology Fragment
Label
Comment O2
Fragment
Label read ontology Comment
query wikipedia titles
Lang=x
Lang=a query wikipedia titles
query translated titles Lang=x
Fragment Label Comment Fragment Label Comment Lang=a
Lang=y Fragment Label Comment Fragment Label Comment
Lang=y query translated titles
Lang=z translated titles translated titles
Lang=z Fragment Label Comment Fragment Label
Comment Lang=y comparison
Lang=z improve the recall measure we have tried another search approach, i.e., splitting each term into individual tokens and searching for those tokens individually. For example, the query for the string Passive conference participant will therefore contain three single searches with passive, conf erence and participant. Both search approaches are shown in Fig. 2 in pseudo code.
Our own tests showed that the individual tokens search (ITS) will result in a better recall, but a lower precision. To have a look at the F-Measure between the two approaches, the first idea of simple search can produce better values. Therefore this approach is was submitted. 1.3
No adaptions for the evaluation are made. } } } } float getsimilarity(term1, term2) { titlesForTerm1 = getAllTitles(term1); titlesForTerm2 = getAllTitles(term2); commonTitles = intersectionOf(titlesForTerm1, titlesForTerm2); allTitles = unionOf(titlesForTerm1, titlesForTerm2); return #(commonTitles) / #(allTitles); List<WikipediaPage> getAllTitles(searchTerm) { removeStopwords(searchTerm); removePunctuation(searchTerm); if(simpleSearch) {
resultList = searchWikipedia(searchTerm); if(individualTokenSearch) { tokens = tokenize(searchTerm); for each token in tokens
resultList = resultList + searchWikipedia(searchTerm); for each page in results
resultList = resultList + getLanguageLinks(page); return resultList;
The WikiMatch tool can be downloaded from http://www.ke.tu-darmstadt. de/resources/ontology-matching/wikimatch.
Since our approach is entirely element-based, removing or replacing labels or comments results in lower F-Mesaure. By removing only one of the describing elements, WikiMatch deals also with the remaining literals and can provide good results. If there are neither labels nor comments, then this approach does not work. On the other hand, removing structural features, such as subclass relations, does not influence the results of WikiMatch. 2.2
The comparison with StringEquiv of the OAEI 2011.5 is not that well, because the recall is not much higher, but therefore the precision is very low (0.997 to 0.864). A nontrivial mapping that is found by our tool is ophthalmic artery and Opthalmic Artery. In the conference track, WikiMatch reached 0.6 F-Measure for ra1. This is better than the baseline2 from OAEI 2011.5. The same applies for ra2. Unfortunately, the conference domain is not well covered in Wikipedia to match special terms like Chair PC and ProgramCommitteeChair. But through the suggestion feature it is possible to find a mapping between Sponsorship and Sponzorship. 2.4
On the Multifarm track, WikiMatch exploits the inter-language links from each returned article. Therefore a mapping between different languages can be found. The best results are achieved for matching English to Spanish (F-Measure 0.29), the worst for ChineseGerman and Chinese-Portuguese (F-Measure 0.1).
The results on the Multifarm track strongly depend on the involved Wikipedia’s sizes, in particular the number of articles and links to other Wikipedias. Fig. 3 depicts the results of WikiMatch in relation to the corresponding Wikipedias’ article counts; Fig. 4 the results in relation to the number of links from the corresponding Wikipedias to other Wikipedias1. It can be observed that the results get better with larger and more strongly inter-linked Wikipedias.
As the number of articles and inter-Wikipedia links grow by around 2% per month (even more rapidly for Chinese, which is currently the smallest and least interlinked Wikipedia used in Multifarm), we expect the results of WikiMatch to improve just by the growth of Wikipedia. The trend lines in Fig. 3 and 4 indicate that about 500,000 additional articles and Wikipedia links lead to an increase of five percentage points in F-Measure. At the current growth rate of Wikipedia, this takes a little less than two years. 2.5
The library track unfortunately did not finish within one week. The reason can be the calculation of the cross product between the concepts of the ontologies, or the generally long times required for looking up concepts in Wikipedia. This requires an more detailed look. 2.6
Like the library track, the ontologies in this track are also too large handle by WikiMatch in its current version. 1 Using numbers obtained from http://stats.wikimedia.org/ 1000000 2000000 3000000 4000000 5000000
H-mean link count Fig. 4. Multifarm results in relation to the corresponding Wikipedias’ inter-wiki link counts
On Multifarm and conference track WikiMatch shows that a simple element based approach can keep up with state of the art tools. Especially using the inter-language links in Wikipedia looks like a promising approach to deal with multi-lingual ontologies. On large tracks the current approach does not scale well and did not finish in time.
In general, like most approaches using web data by querying the web at run-time, WikiMatch is rather slow compared to matchers working entirely internally or only use local resources. 3.2
For improving the approach, we envision to set threshold values dynamically, based on the matched ontologies. In order to cope with the run-time restrictions, it is possible to not use WikiMatch as a single matching approach, but to first match the easy cases (i.e., same or very similar terms) with string-level methods.
At the moment, WikiMatch only uses the page identifiers returned by the search,
ignoring the text snippets, i.e., the portions of the Wikipedia pages that are relevant
for the search term. Using those snippets, e.g., like WeSeE-Match does [
With our work on WikiMatch, we have shown how a large general-purpose resource like Wikipedia can be used for ontology matching. Especially the cross-linking of different language Wikipedias is useful for multi-lingual ontology matching. Furthermore, we have seen that the results of WikiMatch improve with a growing size of Wikipedia – which in turn indicates that the results of WikiMatch will improve in the future merely by the growth of Wikipedia.