If users participate in ontology matching, the goal always is to minimize the amount of necessary interactions while maximizing the gains in alignment quality [2]. Interaction can either happen pre-matching (selection of matching systems or parameter tuning), during the matching process (judging intermediate results or providing sample correspondences), or post-matching (detecting incorrect correspondences and providing missing ones). In this paper, we evaluate an approach that aims to reduce post matching interactions by exploiting concept proximity within ontologies. An initial analysis of reference alignments available for OAEI revealed that, if a correspondence for one element (class or property) of an ontology exists, the probability that a correspondence also exists for a closely connected element is higher than for unconnected elements. Based on this nding, we extracted the closeness criteria depicted in Figure 1. For evaluation, we applied the criteria on candidate alignments that were created by top-performing systems of OAEI 2014 for the anatomy, library, and conference tracks. For each criterion, we determined which elements it would add to the task set, i.e., the selection of ontology elements a user should provide correspondences for. Based on these task sets (U T ) we calculated the expected number of interactions (IE) it would on average take to provide all included correspondences (IC), if elements were presented to the user at random. To assess whether our selection technique is viable, we further compared this value to the amount of interactions it would take users on average to provide the same amount of missing correspondences, if tasks were randomly selected from the entirety of elements that are not included in correspondences after initial, automatic matching.
ProOpbejretyct 2 17
Sub Class 7 7 Domain Class Domain Class Subproperty 10 Superproperty 11 13 := 6 + 7 14 := 7 + 8 15 := 5 + 8
The ratio between the two values is called task set compression. Minimal criteria sets denote the closeness criteria, which yield the corresponding task sets. Since we strive for minimization of user tasks, only the smaller ontology in terms of concept count was considered. As shown in Table 1, an average task set reduction of 60% could be achieved for the conference ontologies of OAEI, while increasing the recall from 0,62 to 0,956. For taxonomy-like ontologies, such as the ones used in the library and anatomy tracks, only marginal compression or even increase in interaction expectancy was achieved. Future work will therefore focus on such cases by nding other, more suitable task selection criteria and adapting existing ones, e.g., by limiting the depth of hierarchy traversal for class relationships. Furthermore, correspondences generated through di erent matcher settings (high precision vs. high recall) could be explored in addition to criteria based solely on ontology structures in order to yield smaller task sets with an increased potential success ratio for user interactions.
ontologies
jUTj IC IE Compression Minimal Criteria Sets Rcand Rcomp minimal criteria sets. The used criteria are numbered according to Figure 1. Rcand is the recall achieved by the automatic matcher, Rcomp the recall after user interaction. of Ontologies. In: Proceedings of the 19th International Conference on Knowledge Engineering and Knowledge Management. pp. 266{281. EKAW '14 (2014)
IEEE Trans. on Knowl. and Data Eng. 25(