Whereas the ontology matching field has developed fully in the last decades, most matching approaches are still limited to generating equivalences between entities of different ontologies. However, for many tasks, finding subsumption relations may be useful. Despite the variety of matching approaches in the literature, most of them rely on string-based techniques as an initial estimate of the likelihood that two elements refer to the same real world phenomenon, hence, the found correspondences represent equivalences with terms similarly written rather than subsumptions. This paper presents an approach relying on background knowledge from BabelNet (BN) [3] and on the notion of context. The latter has been exploited in different ways in ontology matching [2, 4]. They are used for disambiguating the senses that better express the meaning of ontology concepts when looking for subsumption relations between them in BN.
Material and methods. We used the set of 7 ontologies from the OAEI conference data set that are involved in the 21 available reference alignments. In our experiments, compounds with no entry in BN have been pre-processed by removing the modifiers (e.g. “Invited speaker” is a “Speaker”). We empirically selected k=2 for the path length and 0.8 as edit distance threshold. We used as reference the subsumptions inferred from the available equivalence reference alignments, using Hermit and the Alignment API 4.5. As many concepts do not have any super or sub concepts, we considered 2 settings: contexts as introduced above and the whole ontology as context for each concept. The best results, which are reported here, were obtained with the latter.
Results and discussion. Table 1 shows the results (measures were computed using the Alignment API). Overall, the best results are obtained when considering alignments close to those expected (extended and semantic measures) rather than exact ones. Looking at the results for each pair of ontologies, the best results where obtained for different pairs when using the different measures: edas-ekaw (classical), confOf-edas (extended) and conference-sigkdd (semantic). The overall low results are mainly due to two reasons: a high number of concepts can not be found in BN and using the modifier does help so much in this task; the construction of contexts suffers from the lack of annotations in the ontologies (as well many concepts do not have any super or sub concepts), and hence, contexts are not rich enough for disambiguating the synsets.
We presented an approach for generating subsumption correspondences relying on BabelNet. This task is still a gap in the field and the initial results presented here can be improved in different ways. We plan to improve the disambiguation strategy, exploiting word embeddings, to automatically enrich the ontology with annotations, to adopt a hybrid approach combining both lexical and background knowledge, to work on the confidence of the correspondences, and to look for other relations like meronymy.