Domain-aware ontology matching on the semantic web

The rise of the Semantic Web and Linked Open Data has led to a large number of different and heterogeneous datasets and ontologies published on the Web. This kind of heterogeneity of datasets has created a need to interlink them, i.e. to make explicit that two resources of different datasets in fact represent the exact same concept. This is the task commonly known as "ontology matching." This thesis presents work on the development and evaluation of a domain-aware ontology matching approach geared at interlinking domain-specific ontologies to the cross-domain ontology DBpedia. We present a bootstrap method to automatically derive domain knowledge from an initial set of high confidence matches, and show several methods to translate this domain knowledge into an optimized filter that can be used to limit matches to. The domain derivation and optimization is accomplished in a fully unsupervised way. To explore the generalizability of the approach, we perform evaluations on two use cases: Last.fm artists and UMLS medical terms. We show that for ambiguous data such as artist and band names, the proposed approach outperforms a traditional matching approach by as much as 17.1% on an F-score scale of quality. Improvements for highly specialized data such as medical terms are less significant but can be gained by taking special measures. Finally, given an example application of medical training simulators, we show that the domain-aware approach is also potentially very useful when adapted to the semantic enrichment of free text, where a gain of as much as 26.6% is attained over traditional annotation approaches.

Subject

semantic web
linked data
ontology alignment

To reference this document use:

http://resolver.tudelft.nl/uuid:2f98add8-e471-47e8-b1b1-bf93fd76acc5

Embargo date

2012-02-02

Part of collection

Student theses

Document type

master thesis

Rights

Files

PDF

Thesis_-_Kristian_Slabbekoorn.pdf

3.72 MB

Close viewer