Recently, a paper by Aviv Segev and me was accepted to IEEE Transactions on Services Computing. The article, with the rather long name of “Context-Based Matching and Ranking of Web Services for Composition” tries to challenge a simple problem. Given a Web service in WSDL (short documents that describe Web services) and a set of ontology domains, to which ontology domains does this service belongs, and how strong is this “belongness”.
The simple solution is to take the words in the WSDL, to take the words in the ontology, and to compare them. This is how documents are classified to topics in a text corpus. The problem is that the WSDL documents, that describe Web services, contain very little text. At the worst case, they can contain just the names of the input and output parameters. In this paper, we searched for a solution on the Web. We took the words form the WSDL descriptions and thrown them to search engines. We collected words from the results, and ranked these words according to the number of times they appeared for different elements of the service. For example, if a service input parameter was “zip” and the output was “address”, and running these words in a search engine returned words such as “location” (due to both zip and address) and “compression” (due to zip, such as in win-zip), we ranked “location” higher than “compression”. We then took those words, added them to the WSDL words and used all of them in the comparison process.
We had compared our method to simple string matching, to tf/idf and to a number of other methods. We found out that our method and tf/idf composed together outperformed other every single method.
0 Responses to “Categorizing Web Services”