This paper describes the updates in EVOCROS, a crosslingual ontology alignment system suited to create mappings between ontologies described in di erent natural language. Our tool combines syntactic and semantic similarity measures with information retrieval techniques. The semantic similarity is computed via NASARI vectors used together with BabelNet, which is a domain-neutral semantic network. In particular, we investigate the use of rank aggregation techniques in the cross-lingual ontology alignment task. The tool employs automatic translation to a pivot language to consider the similarity. EVOCROS was tested and obtained high quality alignment in the Multifarm dataset. We discuss the experimented con gurations and the achieved results in OAEI 2019. This is our second participation in OAEI.
1.1
EVOCROS is a cross-lingual ontology alignment tool. The newest version of the tool leverages supervised methods of ranking aggregation techniques exploiting labeled information (i.e., training data) and ground-truth relevance to boost the e ectiveness of a new ranker. Our goal is to leverage rank aggregation in cross-lingual mapping, by generating ranked lists based on distinct similarity measurements between the concepts of source and target ontologies. 1.2
The tool is developed in Python 3 and uses learning to rank techniques implemented in the well-known library RankLib. We model the mapping problem as an information retrieval query. Figure 1 depicts the work ow of the proposed technique. The inputs are source and target ontologies written in Web Ontology Language (OWL). These ontologies are converted to objects. The rst step is the pre-processing of the source and target input ontologies, converting them into owlready2 objects. Each concept of the source ontology is compared to all concepts of the target ontology.
RankLib: https://sourceforge.net/p/lemur/wiki/RankLib/ (As of November 16, 2019).
Python 3 library to manipulate ontologies as objects.
Each entity of the source ontology is compared with all entities of the same type found in the target ontology (i.e., classes are matched to classes and properties are matched to properties). In this sense, for each entity ei in the source ontology OX , we calculate the similarity value with each entity ej in the target ontology OY (Figure 2), thus generating a ranked list frank1; rank2; rank3; rank4g for each similarity measure used (cf. Figure 3).
For similarity measures that rely on monolingual comparison (i.e., syntactic and WordNet), the automatic translation of labels of entities ei 2 OX and ej 2 OY to a pivot language is used by leveraging Google Translate API during runtime. These similarity comparisons generate k ranks, each one based on a di erent similarity measure. We use the measures to generate the ranks, thus adding the exibility to the use or the addition of di erent similarity measures without disrupting the technique.
The ranks are then aggregated using LambdaMART [
Alignment results are available at https://github.com/jmdestro/evocros-results
(As of November 16, 2019).
In this section, we describe the results obtained in the experiments conducted
in OAEI 2019.
We consider the MultiFarm dataset [
A subset of all languages was used for training and validation. The subsets are 10% of queries for training set, 15% queries for validation set, and 75% queries for testing. These subsets were generated per language and then combined, so the algorithms were trained, validated and tested using all languages at once. The comparable gold standard (i.e., MultiFarm manually curated mappings) were adjusted to contain only the queries related to the testing subset. In this sense, a lower number of entities was considered in the tests, because we removed the set of queries used in training and validation from the reference mappings to ensure consistency.
Table 1 presents the obtained values for precision, recall, and f-measure for each language pair tested. The precision, recall, and f-measure scores have the same value due to the nature of the experiments. Our approach generates n : n mappings, where n = jOX j = jOY j because the ontologies are translations of each other to di erent natural languages, thus every entity in the source ontology presents a correspondence in the target ontology. In this sense, both the gold standard and the generated mappings have the same size because each query (i.e., each entity in the source ontology) generates a mapping between the query (source entity) and the top-1 result of the nal aggregated rank. Results
In this section, we discuss our results and the ways to improve the system. 3.1
The tool had satisfactory results, with competitive f-measure, but the execution time was exceedingly long due even with local caches for Babelnet NASARI vectors. This is due to the amount of comparisons required during execution because each concept or attribute in the source ontology is compared against all concepts and attributes of the target ontology. 3.2
This was the second evaluation of the system and results are encouraging. Our
main goals for future work are: Reduce execution time: the tool has a long
execution time even with local caches. Our future work will explore ontology
partitioning during the pre-processing stage of the matching task to reduce
the amount of comparisons needed, thus improving the execution time. Bag
of graphs: ontologies can be represented as graphs, thus allowing for
partitioning [
Although we were not participating, our tool was executed on the Knowledge Graph track. There were issues during the evaluation phase, preventing the system to fully participate in both Multifarm and KG tracks. 4
The newest version of EVOCROS proposed an approach considering four similarity measures to build ranks and used a supervised method of rank aggregation. This is the second participation of the system in OAEI. The evaluation with the Multifarm dataset con rmed the quality of mappings generated by our technique. For future work, we plan to improve our cross-lingual alignment proposal by considering di erent combinations of similarity measures and di erent ways of computing the syntactic and semantic similarities taking into account additional stages in the pre-processing of the ontology.
This work was supported by S~ao Paulo Research Foundation (FAPESP): grant #2017/02325-5.