This year out of five matchers registered for the complex track one matcher, AMLC, managed to match the consensus conference dataset. AROA and CANARD need instances for matching which are missing in this dataset. Lily did not generate any complex correspondence.
The following table summarizes numbers of complex equivalence correspondences generated by AMLC and their types per each ontology pair. AMLC focused on three types of complex equivalence correspondences: those with union of the classes, those with attribute occurence restriction and those with attribute domain restriction.
| ontology pair | #UnionOfClasses | #AttributeOccurenceRestriction | #AttributeDomainRestriction | #complex equivalence correspondences |
|---|---|---|---|---|
| conference-ekaw | 2 | 7 | 16 | 25 |
| cmt-conference | 1 | 6 | 4 | 11 |
| cmt-ekaw | 3 | 9 | 5 | 17 |
The following table summarizes the evaluation resuts of AMLC on three ontology pairs.
| ontology pair | P | R | F-measure |
|---|---|---|---|
| conference-ekaw | 0.28 | 0.32 | 0.30 |
| cmt-conference | 0.18 | 0.22 | 0.20 |
| cmt-ekaw | 0.47 | 0.57 | 0.52 |
| average | 0.31 | 0.37 | 0.34 |
Based on the evaluation we have the following findings. (1) Although the performance in terms of precision and recall decreased for AMLC, AMLC managed to find more True Positives. (2) Since AMLC provides confidence, it could be possible to include confidence into the evaluation and this could improve the performance results. (3) AMLC discovered one more kind of complex mappings: the union of classes. (4) Since the evaluation was manual and this kind of evaluation is demanding, we cannot guarantee avoiding errors during the evaluation. This task should be supported by an automatic process.
The generated alignments by AMLC are available