An ontology matching task MT is composed of a pair of ontologies O1 and O2 and possibly an associated reference alignment MRA. The objective of a matching task is to discover an overlapping of O1 and O2 in the form of an alignment M. The size or search space of a matching task is typically bound to the size of the Cartesian product between the entities of the input ontologies. Large-scale ontology matching tasks still pose serious challenges to state-of-the-art ontology alignment systems [2]. In this paper we propose a novel method to effectively divide an input ontology matching task MT into several (independent) and more manageable (sub)tasks. This method relies on an efficient lexical index (as in LogMap [3]), a neural embedding model [4] and locality modules [5]. Unlike other state-of-the-art approaches, our method provides guarantees about the preservation of the coverage of the relevant ontology alignment.
1 Introduction
The approach presented in this paper relies on an ‘inverted’ lexical index (we will refer
to this index as LexI), commonly used in information retrieval applications, and also
used in ontology alignment systems like LogMap [
Creating matching subtasks from LexI
Deriving mappings from LexI. Each entry in LexI, after discarding entries pointing to
only one ontology, is a source of candidate mappings. For instance the example in
Table 1 suggests that there is a (potential) mapping between the entities O1:Serous acinus
and O2:Liver acinus since they are associated to the same entry in LexI facinusg. The
mappings derived from LexI are not necessarily correct but will link lexically related
? An extended version of this paper is available in arXiv.org [
entities. We refer to the set of all mappings suggested by LexI as MLexI. Note that
the input ontologies. Mappings outside MLexI will rarely be discovered by standard
matching systems as they typically rely on lexical similarity measures [
didate mappings MLexI and, as a consequence, to large modules O1LexI and O2LexI. These
modules, although smaller than O1 and O2, can still be challenging for many ontology
matching systems. A solution is to divide the entries in LexI into more than one cluster.
Figure 1 shows an overview of the pipeline where LexI is split into n clusters and these
clusters lead to n matching subtasks DMnT = fMT1LexI; : : : ; MTnLexIg.
Clustering strategies. We have implemented two clustering strategies which we refer to
as: naive and neural embedding. The naive strategy implements a very simple algorithm
that randomly splits the entries in LexI into a given number of clusters of the same
size, while the neural embedding strategy aims at identifying more accurate clusters
and relies on the StarSpace toolkit and its neural embedding model [
to divide an ontology matching task.2 We rely on the datasets of the Ontology Alignment Evaluation Initiative (OAEI), more specifically, on the matching tasks provided in the anatomy (AMA-NCIA), largebio (FMA-NCI, FMA-SNOMED, SNOMED-NCI) and phenotype (HPO-MP, DOID-ORDO) tracks.
Adequacy of clustering strategies. We have evaluated the clustering strategies in terms of the coverage with respect to the OAEI reference alignments3 and the size of the ontologies in the matching subtasks. We have compared the two strategies for different number of matching subtasks n 2 f2; 5; 10; 20; 50; 100; 200g. Figure 2 shows the coverage of the different divisions DMnT for the naive (left) and neural embedding (right) strategies. The results are very good and, in the worst case, approx. 93% of the available 200 . reference mappings in SNOMED-NCI are covered by the matching subtasks in DMT
The scatter plots in Figure 3 visualize, for the FMA-NCI case, the size of the source
modules against the size of the target modules for the matching subtasks in each
division DMnT . For instance, the (orange) triangles represent points jSig(O1i)j; jSig(O2i)j
being O1i and O2i the ontologies (with i=1,. . . ,5) in the matching subtasks of DnM5T . The
naive strategy leads to rather balanced and similar tasks for each division DMT , while
the neural embedding strategy has more variability in the size of the tasks within a given
division DMnT . Nonetheless, on average, the size of the matching subtasks in the neural
embedding strategy are significantly smaller than the ones in the naive strategy.
Evaluation of OAEI systems. We have selected the following four systems from the
latest OAEI campaigns: Mamba, FCA-Map, KEPLER, and POMap. These systems were
unable to complete, given some computational constraints, some OAEI tasks. Table 2
n
shows the obtained results with different divisions DMT computed by the naive and
2 Extended evaluation material in [
Fig. 3: Source and target module sizes in the computed subtasks for FMA-NCI.
neural embedding strategies, for the OAEI tasks that the selected systems failed to compute results. For example, Mamba was able to complete the OAEI 2015 Anatomy track with divisions DM20 T and DM50 T involving 20 and 50 matching subtasks, respectively.