Abstract. The Ontology Alignment Evaluation Initiative (OAEI) aims at comparing ontology matching systems on precisely defined test cases. These test cases can be based on ontologies of different levels of complexity (from simple thesauri to expressive OWL ontologies) and use different evaluation modalities (e.g., blind evaluation, open evaluation, or consensus). The OAEI 2019 campaign offered 11 tracks with 29 test cases, and was attended by 20 participants. This paper is an overall presentation of that campaign.
The Ontology Alignment Evaluation Initiative1 (OAEI) is a coordinated international initiative, which organizes the evaluation of an increasing number of ontology matching systems [21, 23]. The main goal of the OAEI is to compare systems and algorithms openly and on the same basis, in order to allow anyone to draw conclusions about the best matching strategies. Furthermore, our ambition is that, from such evaluations, developers can improve their systems.
Two first events were organized in 2004: (i) the Information Interpretation and
Integration Conference (I3CON) held at the NIST Performance Metrics for Intelligent
Systems (PerMIS) workshop and (ii) the Ontology Alignment Contest held at the
Evaluation of Ontology-based Tools (EON) workshop of the annual International Semantic
Web Conference (ISWC) [48]. Then, a unique OAEI campaign occurred in 2005 at the
workshop on Integrating Ontologies held in conjunction with the International
Conference on Knowledge Capture (K-Cap) [5]. From 2006 until the present, the OAEI
campaigns were held at the Ontology Matching workshop, collocated with ISWC [
Since 2011, we have been using an environment for automatically processing evaluations (x2.1) which was developed within the SEALS (Semantic Evaluation At Large Scale) project3. SEALS provided a software infrastructure for automatically executing evaluations and evaluation campaigns for typical semantic web tools, including ontology matching. Since OAEI 2017, a novel evaluation environment called HOBBIT (x2.1) was adopted for the HOBBIT Link Discovery track, and later extended to enable the evaluation of other tracks. Some tracks are run exclusively through SEALS and others through HOBBIT, but several allow participants to choose the platform they prefer.
This paper synthesizes the 2019 evaluation campaign and introduces the results provided in the papers of the participants. The remainder of the paper is organized as follows: in x2, we present the overall evaluation methodology; in x3 we present the tracks and datasets; in x4 we present and discuss the results; and finally, x5 discusses the lessons learned. 1 http://oaei.ontologymatching.org 2 http://om2019.ontologymatching.org 3 http://www.seals-project.eu
The OAEI evaluation was carried out in one of two alternative platforms: the SEALS client or the HOBBIT platform. Both have the goal of ensuring reproducibility and comparability of the results across matching systems.
The SEALS client was developed in 2011. It is a Java-based command line interface for ontology matching evaluation, which requires system developers to implement a simple interface and to wrap their tools in a predefined way including all required libraries and resources. A tutorial for tool wrapping is provided to the participants, describing how to wrap a tool and how to run a full evaluation locally.
The HOBBIT platform4 was introduced in 2017. It is a web interface for linked data and ontology matching evaluation, which requires systems to be wrapped inside docker containers and includes a SystemAdapter class, then being uploaded into the HOBBIT platform [34].
Both platforms compute the standard evaluation metrics against the reference alignments: precision, recall and F-measure. In test cases where different evaluation modalities are required, evaluation was carried out a posteriori, using the alignments produced by the matching systems. 2.2
As in previous years, the OAEI 2019 campaign was divided into three phases: preparatory, execution, and evaluation.
In the preparatory phase, the test cases were provided to participants in an initial assessment period between June 15th and July 15th, 2019. The goal of this phase is to ensure that the test cases make sense to participants, and give them the opportunity to provide feedback to organizers on the test case as well as potentially report errors. At the end of this phase, the final test base was frozen and released.
During the ensuing execution phase, participants test and potentially develop their matching systems to automatically match the test cases. Participants can self-evaluate their results either by comparing their output with the reference alignments or by using either of the evaluation platforms. They can tune their systems with respect to the nonblind evaluation as long as they respect the rules of the OAEI. Participants were required to register their systems and make a preliminary evaluation by July 31st. The execution phase was terminated on September 30th, 2019, at which date participants had to submit the (near) final versions of their systems (SEALS-wrapped and/or HOBBIT-wrapped).
During the evaluation phase, systems were evaluated by all track organizers. In case minor problems were found during the initial stages of this phase, they were reported to the developers, who were given the opportunity to fix and resubmit their systems. Initial results were provided directly to the participants, whereas final results for most tracks were published on the respective OAEI web pages by October 14th, 2019. 4 https://project-hobbit.eu/outcomes/hobbit-platform/
This year’s OAEI campaign consisted of 11 tracks gathering 29 test cases, all of which were based on OWL ontologies. They can be grouped into: – Schema matching tracks, which have as objective matching ontology classes and/or properties. – Instance Matching tracks, which have as objective matching ontology instances. – Instance and Schema Matching tracks, which involve both of the above. – Complex Matching tracks, which have as objective finding complex correspondences between ontology entities. – Interactive tracks, which simulate user interaction to enable the benchmarking of interactive matching algorithms.
The tracks are summarized in Table 1.
SEALS SEALS SEALS SEALS
both SEALS HOBBIT HOBBIT SEALS SEALS
SEALS = =, <= =, <= = = Open evaluation is made with already published reference alignments and blind evaluation is made by organizers, either from reference alignments unknown to the participants or manually. 3.1 The anatomy track comprises a single test case consisting of matching two fragments of biomedical ontologies which describe the human anatomy5 (3304 classes) and the anatomy of the mouse6 (2744 classes). The evaluation is based on a manually curated reference alignment. This dataset has been used since 2007 with some improvements over the years [15].
Systems are evaluated with the standard parameters of precision, recall, F-measure. Additionally, recall+ is computed by excluding trivial correspondences (i.e., correspondences that have the same normalized label). Alignments are also checked for coherence using the Pellet reasoner. The evaluation was carried out on a server with a 6 core CPU @ 3.46 GHz with 8GB allocated RAM, using the SEALS client. However, the evaluation parameters were computed a posteriori, after removing from the alignments produced by the systems, correspondences expressing relations other than equivalence, as well as trivial correspondences in the oboInOwl namespace (e.g., oboInOwl#Synonym = oboInOwl#Synonym). The results obtained with the SEALS client vary in some cases by 0.5% compared to the results presented below. 3.2
The second edition of biodiversity track features two test cases based on highly overlapping ontologies that are particularly useful for biodiversity and ecology research: matching Environment Ontology (ENVO) to Semantic Web for Earth and Environment Technology Ontology (SWEET), and matching Flora Phenotype Ontology (FLOPO) to Plant Trait Ontology (PTO). The track was motivated by two projects, namely GFBio7 (The German Federation for Biological Data) and AquaDiva8, which aim at providing semantically enriched data management solutions for data capture, annotation, indexing and search [35, 37]. Table 2 summarizes the versions and the sizes of the ontologies used in OAEI 2019. Compared to the first edition, the number of concepts of the ENVO and FOLPO ontologies has increased, which required the creation of new reference alignments for both tasks. 5 www.cancer.gov/cancertopics/cancerlibrary/terminologyresources 6 http://www.informatics.jax.org/searches/AMA_form.shtml 7 www.gfbio.org 8 www.aquadiva.uni-jena.de
To this end, we updated the reference alignments for the two test cases following the same procedure as in the first edition. In particular, alignment files were produced through a hybrid approach consisting of (1) an updated consensus alignment based on matching systems output, then (2) manually validating a subset of unique mappings produced by each system (and adding them to the consensus if considered correct), and finally (3) adding a set of manually generated correspondences. The matching systems used to generate the consensus alignments were those participating in this track last year [4], namely: AML, Lily, LogMap family, POMAP and XMAP.
The evaluation was carried out on a Windows 10 (64-bit) desktop with an Intel Core i5-7500 CPU @ 3.40GHz x 4 with 15.7 Gb RAM allocated, using the SEALS client. Systems were evaluated using the standard metrics. 3.3
The conference track features a single test case that is a suite of 21 matching tasks corresponding to the pairwise combination of 7 moderately expressive ontologies describing the domain of organizing conferences. The dataset and its usage are described in [52].
The track uses several reference alignments for evaluation: the old (and not fully complete) manually curated open reference alignment, ra1; an extended, also manually curated version of this alignment, ra2; a version of the latter corrected to resolve violations of conservativity, rar2; and an uncertain version of ra1 produced through crowd-sourcing, where the score of each correspondence is the fraction of people in the evaluation group that agree with the correspondence. The latter reference was used in two evaluation modalities: discrete and continuous evaluation. In the former, correspondences in the uncertain reference alignment with a score of at least 0.5 are treated as correct whereas those with lower score are treated as incorrect, and standard evaluation parameters are used to evaluated systems. In the latter, weighted precision, recall and F-measure values are computed by taking into consideration the actual scores of the uncertain reference, as well as the scores generated by the matching system. For the sharp reference alignments (ra1, ra2 and rar2), the evaluation is based on the standard parameters, as well the F0:5-measure and F2-measure and on conservativity and consistency violations. Whereas F1 is the harmonic mean of precision and recall where both receive equal weight, F2 gives higher weight to recall than precision and F0:5 gives higher weight to precision higher than recall.
Two baseline matchers are used to benchmark the systems: edna string edit distance matcher; and StringEquiv string equivalence matcher as in the anatomy test case.
The evaluation was carried out on a Windows 10 (64-bit) desktop with an Intel Core i7–8550U (1,8 GHz, TB 4 GHz) x 4 with 16 GB RAM allocated using the SEALS client. Systems were evaluated using the standard metrics. 3.4
The Disease and Phenotype is organized by the Pistoia Alliance Ontologies Mapping project team9. It comprises 2 test cases that involve 4 biomedical ontologies covering the disease and phenotype domains: Human Phenotype Ontology (HP) versus 9 http://www.pistoiaalliance.org/projects/ontologies-mapping/ Mammalian Phenotype Ontology (MP) and Human Disease Ontology (DOID) versus Orphanet and Rare Diseases Ontology (ORDO). Currently, correspondences between these ontologies are mostly curated by bioinformatics and disease experts who would benefit from automation of their workflows supported by implementation of ontology matching algorithms. More details about the Pistoia Alliance Ontologies Mapping project and the OAEI evaluation are available in [25]. Table 3.4 summarizes the versions of the ontologies used in OAEI 2019.
The reference alignments used in this track are silver standard consensus alignments automatically built by merging/voting the outputs of the participating systems in 2016, 2017 and 2018 (with vote=3). Note that systems participating with different variants and in different years only contributed once in the voting, that is, the voting was done by family of systems/variants rather than by individual systems. The HP-MP silver standard thus produced contains 2232 correspondences, whereas the DOID-ORDO one contains 2808 correspondences.
Systems were evaluated using the standard parameters as well as the number of unsatisfiable classes computed using the OWL 2 reasoner HermiT [41]. The evaluation was carried out in a Ubuntu 18 Laptop with an Intel Core i5-6300HQ CPU @ 2.30GHz x 4 and allocating 15 Gb of RAM. 3.5
The large biomedical ontologies (largebio) track aims at finding alignments between the large and semantically rich biomedical ontologies FMA, SNOMED-CT, and NCI, which contain 78,989, 306,591 and 66,724 classes, respectively. The track consists of six test cases corresponding to three matching problems (FMA-NCI, FMA-SNOMED and SNOMED-NCI) in two modalities: small overlapping fragments and whole ontologies (FMA and NCI) or large fragments (SNOMED-CT).
The reference alignments used in this track are derived directly from the UMLS Metathesaurus [6] as detailed in [32], then automatically repaired to ensure logical coherence. However, rather than use a standard repair procedure of removing problem causing correspondences, we set the relation of such correspondences to “?” (unknown). These “?” correspondences are neither considered positive nor negative when evaluating matching systems, but are simply ignored. This way, systems that do not perform alignment repair are not penalized for finding correspondences that (despite causing incoherences) may or may not be correct, and systems that do perform alignment repair are not penalized for removing such correspondences. To avoid any bias, correspondences were considered problem causing if they were selected for removal by any of the three established repair algorithms: Alcomo [39], LogMap [31], or AML [43]. The reference alignments are summarized in Table 4.
The evaluation was carried out in a Ubuntu 18 Laptop with an Intel Core i5-6300HQ CPU @ 2.30GHz x 4 and allocating 15 Gb of RAM. Evaluation was based on the standard parameters (modified to account for the “?” relations) as well as the number of unsatisfiable classes and the ratio of unsatisfiable classes with respect to the size of the union of the input ontologies. Unsatisfiable classes were computed using the OWL 2 reasoner HermiT [41], or, in the cases in which HermiT could not cope with the input ontologies and the alignments (in less than 2 hours) a lower bound on the number of unsatisfiable classes (indicated by ) was computed using the OWL2 EL reasoner ELK [36]. 3.6
The multifarm track [40] aims at evaluating the ability of matching systems to deal with ontologies in different natural languages. This dataset results from the translation of 7 ontologies from the conference track (cmt, conference, confOf, iasted, sigkdd, ekaw and edas) into 10 languages: Arabic (ar), Chinese (cn), Czech (cz), Dutch (nl), French (fr), German (de), Italian (it), Portuguese (pt), Russian (ru), and Spanish (es). The dataset is composed of 55 pairs of languages, with 49 matching tasks for each of them, taking into account the alignment direction (e.g. cmten !edasde and cmtde !edasen are distinct matching tasks). While part of the dataset is openly available, all matching tasks involving the edas and ekaw ontologies (resulting in 55 24 matching tasks) are used for blind evaluation.
We consider two test cases: i) those tasks where two different ontologies (cmt!edas, for instance) have been translated into two different languages; and ii) those tasks where the same ontology (cmt!cmt) has been translated into two different languages. For the tasks of type ii), good results are not only related to the use of specific techniques for dealing with cross-lingual ontologies, but also on the ability to exploit the identical structure of the ontologies.
The reference alignments used in this track derive directly from the manually curated Conference ra1 reference alignments. The systems have been executed on a Ubuntu Linux machine configured with 8GB of RAM running under a Intel Core CPU 2.00GHz x4 processors, using the SEALS client. 3.7 The Link Discovery track features two test cases, Linking and Spatial, that deal with link discovery for spatial data represented as trajectories i.e., sequences of longitude, latitude pairs. The track is based on two datasets generated from TomTom10 and Spaten [12].
The Linking test case aims at testing the performance of instance matching tools that implement mostly string-based approaches for identifying matching entities. It can be used not only by instance matching tools, but also by SPARQL engines that deal with query answering over geospatial data. The test case was based on SPIMBENCH [44], but since the ontologies used to represent trajectories are fairly simple and do not consider complex RDF or OWL schema constructs already supported by SPIMBENCH, only a subset of the transformations implemented by SPIMBENCH was used. The transformations implemented in the test case were (i) string-based with different (a) levels, (b) types of spatial object representations and (c) types of date representations, and (ii) schema-based, i.e., addition and deletion of ontology (schema) properties. These transformations were implemented in the TomTom dataset. In a nutshell, instance matching systems are expected to determine whether two traces with their points annotated with place names designate the same trajectory. In order to evaluate the systems a ground truth was built that contains the set of expected links where an instance s1 in the source dataset is associated with an instance t1 in the target dataset that has been generated as a modified description of s1.
The Spatial test case aims at testing the performance of systems that deal with topological relations proposed in the state of the art DE-9IM (Dimensionally Extended nine-Intersection Model) model [47]. The benchmark generator behind this test case implements all topological relations of DE-9IM between trajectories in the two dimensional space. To the best of our knowledge such a generic benchmark, that takes as input trajectories and checks the performance of linking systems for spatial data does not exist. The focus for the design was (a) on the correct implementation of all the topological relations of the DE-9IM topological model and (b) on producing datasets large enough to stress the systems under test. The supported relations are: Equals, Disjoint, Touches, Contains/Within, Covers/CoveredBy, Intersects, Crosses, Overlaps. The test case comprises tasks for all the DE-9IM relations and for LineString/LineString and LineString/Polygon cases, for both TomTom and Spaten datasets, ranging from 200 to 2K instances. We did not exceed 64 KB per instance due to a limitation of the Silk system11, in order to enable a fair comparison of the systems participating in this track.
The evaluation for both test cases was carried out using the HOBBIT platform. 3.8
The SPIMBENCH track consists of matching instances that are found to refer to the same real-world entity corresponding to a creative work (that can be a news item, 10 https://www.tomtom.com/en_gr/ 11 https://github.com/silk-framework/silk/issues/57 blog post or programme). The datasets were generated and transformed using SPIMBENCH [44] by altering a set of original linked data through value-based, structurebased, and semantics-aware transformations (simple combination of transformations). They share almost the same ontology (with some differences in property level, due to the structure-based transformations), which describes instances using 22 classes, 31 data properties, and 85 object properties. Participants are requested to produce a set of correspondences between the pairs of matching instances from the source and target datasets that are found to refer to the same real-world entity. An instance in the source dataset can have none or one matching counterpart in the target dataset. The SPIMBENCH task uses two sets of datasets12 with different scales (i.e., number of instances to match): – Sandbox (380 INSTANCES, 10000 TRIPLES). It contains two datasets called source (Tbox1) and target (Tbox2) as well as the set of expected correspondences (i.e., reference alignment). – Mainbox (1800 CWs, 50000 TRIPLES). It contains two datasets called source (Tbox1) and target (Tbox2). This test case is blind, meaning that the reference alignment is not given to the participants.
In both cases, the goal is to discover the correspondences among the instances in the source dataset (Tbox1) and the instances in the target dataset (Tbox2).
The evaluation was carried out using the HOBBIT platform. 3.9
The Knowledge Graph track was run for the second year. The task of the track is to match pairs of knowledge graphs, whose schema and instances have to be matched simultaneously. The individual knowledge graphs are created by running the DBpedia extraction framework on eight different Wikis from the Fandom Wiki hosting platform13 in the course of the DBkWik project [27, 26]. They cover different topics (movies, games, comics and books) and three Knowledge Graph clusters shares the same domain e.g. star trek, as shown in Table 5.
The evaluation is based on reference correspondences at both schema and instance levels. While the schema level correspondences were created by experts, the instance correspondences were extracted from the wiki page itself. Due to the fact that not all inter wiki links on a page represent the same concept a few restrictions were made: 1) Only links in sections with a header containing “link” are used 2) all links are removed where the source page links to more than one concept in another wiki (ensures the alignments are functional) 3) multiple links which point to the same concept are also removed (ensures injectivity). Since we do not have a correspondence for each instance, class, and property in the graphs, this gold standard is only a partial gold standard.
The evaluation was executed on a virtual machine (VM) with 32GB of RAM and 16 vCPUs (2.4 GHz), with Debian 9 operating system and Openjdk version 1.8.0 212, using the SEALS client (version 7.0.5). We used the -o option in SEALS to provide the 12 Although the files are called Tbox1 and Tbox2, they actually contain a Tbox and an Abox. 13 https://www.wikia.com/ two knowledge graphs which should be matched. We used local files rather than HTTP URLs to circumvent the overhead of downloading the knowledge graphs. We could not use the ”-x” option of SEALS because the evaluation routine needed to be changed for two reasons: first, to differentiate between results for class, property, and instance correspondences, and second, to deal with the partial nature of the gold standard.
The alignments were evaluated based on precision, recall, and f-measure for classes, properties, and instances (each in isolation). The partial gold standard contained 1:1 correspondences and we further assume that in each knowledge graph, only one representation of the concept exists. This means that if we have a correspondence in our gold standard, we count a correspondence to a different concept as a false positive. The count of false negatives is only increased if we have a 1:1 correspondence and it is not found by a matcher. The whole source code for generating the evaluation results is also available14.
As a baseline, we employed two simple string matching approaches. The source code for these matchers is publicly available15.
The interactive matching track aims to assess the performance of semi-automated matching systems by simulating user interaction [42, 14, 38]. The evaluation thus focuses on how interaction with the user improves the matching results. Currently, this track does not evaluate the user experience or the user interfaces of the systems [29, 14].
The interactive matching track is based on the datasets from the Anatomy and Conference tracks, which have been previously described. It relies on the SEALS client’s Oracle class to simulate user interactions. An interactive matching system can present a collection of correspondences simultaneously to the oracle, which will tell the system whether that correspondence is correct or not. If a system presents up to three correspondences together and each correspondence presented has a mapped entity (i.e., class 14 http://oaei.ontologymatching.org/2019/results/knowledgegraph/ matching-eval-trackspecific.zip 15 http://oaei.ontologymatching.org/2019/results/knowledgegraph/ kgBaselineMatchers.zip or property) in common with at least one other correspondence presented, the oracle counts this as a single interaction, under the rationale that this corresponds to a scenario where a user is asked to choose between conflicting candidate correspondences. To simulate the possibility of user errors, the oracle can be set to reply with a given error probability (randomly, from a uniform distribution). We evaluated systems with four different error rates: 0.0 (perfect user), 0.1, 0.2, and 0.3.
In addition to the standard evaluation parameters, we also compute the number of requests made by the system, the total number of distinct correspondences asked, the number of positive and negative answers from the oracle, the performance of the system according to the oracle (to assess the impact of the oracle errors on the system) and finally, the performance of the oracle itself (to assess how erroneous it was).
The evaluation was carried out on a server with 3.46 GHz (6 cores) and 8GB RAM allocated to the matching systems. Each system was run ten times and the final result of a system for each error rate represents the average of these runs. For the Conference dataset with the ra1 alignment, precision and recall correspond to the micro-average over all ontology pairs, whereas the number of interactions is the total number of interactions for all the pairs. 3.11
The complex matching track is meant to evaluate the matchers based on their ability to generate complex alignments. A complex alignment is composed of complex correspondences typically involving more than two ontology entities, such as o1:AcceptedPaper o2:Paper u o2:hasDecision.o2:Acceptance. Four datasets with their own evaluation process have been proposed [51].
The complex conference dataset is composed of three ontologies: cmt, conference and ekaw from the conference dataset. The reference alignment was created as a consensus between experts. In the evaluation process, the matchers can take the simple reference alignment ra1 as input. The precision and recall measures are manually calculated over the complex equivalence correspondences only.
The populated complex conference is a populated version of the Conference dataset. 5 ontologies have been populated with more or less common instances resulting in 6 datasets (6 versions on the seals repository: v0, v20, v40, v60, v80 and v100). The alignments were evaluated based on Competency Questions for Alignment, i.e., basic queries that the alignment should be able to cover [49]. The queries are automatically rewritten using 2 systems: that from [50] which covers (1:n) correspondences with EDOAL expressions; and a system which compares the answers (sets of instances or sets of pairs of instances) of the source query and the source member of the correspondences and which outputs the target member if both sets are identical. The best rewritten query scores are kept. A precision score is given by comparing the instances described by the source and target members of the correspondences.
The Hydrography dataset consists of matching four different source ontologies (hydro3, hydrOntology-translated, hydrOntology-native, and cree) to a single target ontology (SWO) [9]. The evaluation process is based on three subtasks: given an entity from the source ontology, identify all related entities in the source and target ontology; given an entity in the source ontology and the set of related entities, identify the logical relation that holds between them; identify the full complex correspondences. The three subtasks were evaluated based on relaxed precision and recall [16].
The GeoLink dataset derives from the homonymous project, funded under the U.S. National Science Foundation’s EarthCube initiative. It is composed of two ontologies: the GeoLink Base Ontology (GBO) and the GeoLink Modular Ontology (GMO). The GeoLink project is a real-world use case of ontologies, and the instance data is also available and populated into the benchmark. The alignment between the two ontologies was developed in consultation with domain experts from several geoscience research institutions. More detailed information on this benchmark can be found in [54, 55]. Evaluation was done in the same way as with the Hydrography dataset. The evaluation platform was a MacBook Pro with a 2.5 GHz Intel Core i7 processor and 16 GB of 1600 MHz DDR3 RAM running mac OS Yosemite version 10.10.5.
The Taxon dataset is composed of four knowledge bases containing knowledge about plant taxonomy: AgronomicTaxon, AGROVOC, TAXREF-LD and DBpedia. The evaluation is two-fold: first, the precision of the output alignment is manually assessed; then, a set of source queries are rewritten using the output alignment. The rewritten target query is then manually classified as correct or incorrect. A source query is considered successfully rewritten if at least one of the target queries is semantically equivalent to it. The proportion of source queries successfully rewritten is then calculated (QWR in the results table). The evaluation over this dataset is open to all matching systems (simple or complex) but some queries can not be rewritten without complex correspondences. The evaluation was performed with an Ubuntu 16.04 machine configured with 16GB of RAM running under a i7-4790K CPU 4.00GHz x 8 processors. 4
4.1
Following an initial period of growth, the number of OAEI participants has remained approximately constant since 2012, which is slightly over 20. This year we count with 20 participating systems. Table 6 lists the participants and the tracks in which they competed. Some matching systems participated with different variants (AML, LogMap) whereas others were evaluated with different configurations, as requested by developers (see test case sections for details).
A number of participating systems use external sources of background knowledge, which are especially critical in matching ontologies in the biomedical domain. LogMapBio uses BioPortal as mediating ontology provider, that is, it retrieves from BioPortal the most suitable top-10 ontologies for each matching task. LogMap uses normalizations and spelling variants from the general (biomedical) purpose SPECIALIST Lexicon. AML has three sources of background knowledge which can be used as mediators between the input ontologies: the Uber Anatomy Ontology (Uberon), the Human Disease Ontology (DOID) and the Medical Subject Headings (MeSH). XMAP and Lily use a dictionary of synonyms (pre)extracted from the UMLS Metathesaurus. In addition Lily also uses a dictionary of synonyms (pre)extracted from BioPortal. Confidence pertains to the confidence scores returned by the system, with X indicating that they are non-boolean; # indicates that the system did not participate in the track; indicates that it participated fully in the track; and G# indicates that it participated in or completed only part of the tasks of the track. The results for the Anatomy track are shown in Table 7. Of the 12 systems participating in the Anatomy track, 10 achieved an F-measure higher than the StringEquiv baseline. Two systems were first time participants (Wiktionary and AGM). Long-term participating systems showed few changes in comparison with previous years with respect to alignment quality (precision, recall, F-measure, and recall+), size and run time. The exceptions were LogMapBio which increased in both recall+ (from 0.756 to 0.801) and alignment size (by 57 correspondences) since last year, and ALIN that increased in F-measure (from 0.758 to 0.813) and recall+ (from 0.0 to 0.365), as well as had a substantial increase of 158 correspondences since last year.
In terms of run time, 5 out of 12 systems computed an alignment in less than 100 seconds, a ratio which is similar to 2018 (6 out of 14). LogMapLite remains the system with the shortest runtime. Regarding quality, AML remains the system with the highest F-measure (0.943) and recall+ (0.832), but 3 other systems obtained an F-measure above 0.88 (LogMapBio, POMap++, and LogMap) which is at least as good as the best systems in OAEI 2007-2010. Like in previous years, there is no significant correlation between the quality of the generated alignment and the run time. Four systems produced coherent alignments. Five of the systems participating this year had participated in this track in OAEI 2018: AML, LogMap family systems (LogMap, LogMapBio and LogMapLT) and POMAP. Three were new participants: DOME, FCAMapKG and LogMapKG. The newcomers DOME, FCAMapKG did not register explicitly to this track but could cope with at least one task so we did include their results.
We observed a slight increase in the number of systems (8 systems) that succeeded to generate alignments for the FLOPO-PTO task in comparison to previous year (7 systems). However, we witnessed a slight decrease in the number of systems (6 systems) that succeeded to generate alignments for the test ENVO-SWEET in comparison to previous year (7 systems). Lily did not manage to generate mappings for both tasks and LogMapBio did not manage to generated mappings for the ENVO-SWEET task.
As in the previous edition, we used precision, recall and F-measure to evaluate the performance of the participating systems. This year we included the execution times. The results for the Biodiversity and Ecology track are shown in Table 8.
Overall, the results of the participating systems have decreased in terms of Fmeasure for both tasks compared to last year. In terms of run time, most of the systems (except POMAP) computed an alignment in less than 100 seconds.
For the FLOPO-PTO task, AML and LogMapKG achieved the highest F-measure (0.78), with a slight difference in favor of AML. However, AML showed a remarkable decrease in terms of precision (from 0.88 to 0.76) and F-measure (from 0.86 to 0.78) compared to last year. LogMap also showed a slight decrease in terms of F-measure (from 0.80 to 0.78). The DOME system (newcomer) achieved the highest precision (0.99) with quite a good F-measure (0.739).
Regarding the ENVO-SWEET task, AML ranked first in terms of F-measure (0.80), followed by POMAP (0.69), FCAMapKG (0.63) and LogMapKG (0.63). As last year AML showed a very high recall and significant larger alignment than the other top systems, but a comparably lower precision and a slight decrease in terms of F-measure (from 0.84 to 0.80). POMAP ranked second this year with a remarkable decrease in terms of precision (from 0.83 to 0.68) and F-measure (from 0.78 to 0.69). FCAMapKG and LogMapKG showed the highest results in terms of precision (0.80).
AML generated a significantly large number of mappings (much bigger than the size of the reference alignments for both tasks), those alignments were mostly subsumption mappings. In order to evaluate the precision in a more significant manner, we had to calculate an approximation by assessing manually a subset of mappings not present in the reference alignment (around a 100 for each task).
Overall, in this second evaluation, the results obtained from participating systems remained similar with a slight decrease in terms of F-measure compared to last year. It is worth noting that most of the participating systems, and all of the most successful ones use external resources as background knowledge. The conference evaluation results using the sharp reference alignment rar2 are shown in Table 9. For the sake of brevity, only results with this reference alignment and considering both classes and properties are shown. For more detailed evaluation results, please check conference track’s web page.
With regard to two baselines we can group tools according to matcher’s position: four matching systems outperformed both baselines (SANOM, AML, LogMap and Wiktionary); two performed the same as the edna baseline (DOME and LogMapLt); one performed slightly worse than this baseline (ALIN); and two (Lily and ONTMAT1) performed worse than both baselines. Three matchers (ONTMAT1, ALIN and Lily) do
Prec. F0:5-m. F1-m. F2-m. Rec. Inc.Align. Conser.V. Consist.V. not match properties at all. Naturally, this has a negative effect on their overall performance.
The performance of all matching systems regarding their precision, recall and F1measure is plotted in Figure 1. Systems are represented as squares or triangles, whereas the baselines are represented as circles.
With respect to logical coherence [45, 46], only three tools (ALIN, AML and LogMap) have no consistency principle violation (the same tools as last year). This year all tools have some conservativity principle violations as the last year). We should note that these conservativity principle violations can be “false positives” since the entailment in the aligned ontology can be correct although it was not derivable in the single input ontologies.
This year we additionally analyzed the False Positives, i.e. correspondences discovered by the tools which were evaluated as incorrect. The list of the False Positives is available on the conference track’s web page. We looked at the reasons why a correspondence was incorrect or why it was discovered from a general point of view, and defined 3 reasons why alignments are incorrect and 5 reasons why they could have been chosen. Looking at the results, it can be said that when the reason a correspondence was discovered was the same name, all or at least most tools generated the correspondence. False Positives not discovered based on the same name or synonyms were produced by Lily, ONTMAT1 and SANOM. SANOM was the only tool which produced these correspondences based on similar strings. In three cases, a class was matched with a property by DOME (1x), LogMapLt (1x) and Wiktionary (3x).
The Conference evaluation results using the uncertain reference alignments are presented in Table 10.
Out of the 9 alignment systems, five (ALIN, DOME, LogMapLt, ONTMAT1, SANOM) use 1.0 as the confidence value for all matches they identify. The remaining
F1-measure=0.7
F1-measure=0.6 F1-measure=0.5 AML DOME Lily LogMap LogMapLt ONTMAT1 SANOM Wiktionary edna StringEquiv rec=1.0 rec=.8 rec=.6 pre=.6 pre=.8 pre=1.0 Fig. 1. Precision/recall triangular graph for the conference test case. Dotted lines depict level of precision/recall while values of F1-measure are depicted by areas bordered by corresponding lines F1-measure=0.[5j6j7]. four systems (AML, Lily, LogMap, Wiktionary) have a wide variation of confidence values.
When comparing the performance of the matchers on the uncertain reference alignments versus that on the sharp version (with the corresponding ra1), we see that in the discrete case all matchers except Lily performed the same or better in terms of F-measure (Lily’s F-measure dropped almost to 0). Changes in F-measure of discrete cases ranged from -1 to 17 percent over the sharp reference alignment. This was predominantly driven by increased recall, which is a result of the presence of fewer ’controversial’ matches in the uncertain version of the reference alignment.
The performance of the matchers with confidence values always 1.0 is very similar regardless of whether a discrete or continuous evaluation methodology is used, because many of the matches they find are the ones that the experts had high agreement about, while the ones they missed were the more controversial matches. AML produces a fairly wide range of confidence values and has the highest F-measure under both the continuous and discrete evaluation methodologies, indicating that this system’s confidence evaluation does a good job of reflecting cohesion among experts on this task. Of the remaining systems, three (DOME, LogMap, SANOM) have relatively small drops in F-measure when moving from discrete to continuous evaluation. Lily’s performance drops drastically under the discrete and continuous evaluation methodologies. This is because the matcher assigns low confidence values to some matches in which the labels are equivalent strings, which many crowdsourcers agreed with unless there was a compelling technical reason not to. This hurts recall significantly.
Overall, in comparison with last year, the F-measures of most returning matching systems essentially held constant when evaluated against the uncertain reference alignments. The exception was Lily, whose performance in the discrete case decreased dramatically. ONTMAT1 and Wiktionary are two new systems participating in this year. ONTMAT1’s performance in both discrete and continuous cases increases 16 percent in terms of F-measure over the sharp reference alignment from 0.55 to 0.64, which it is mainly driven by increased recall. Wiktionary assigns confidence value of 1.0 to the entities with identical strings in two ontologies, while gives confidence value of 0.5 to other possible candidates. From the results, its performance improves significantly from sharp to discrete and continuous cases. In the OAEI 2019 phenotype track 8 systems were able to complete at least one of the tasks with a 6 hours timeout. Table 11 shows the evaluation results in the HP-MP and DOID-ORDO matching tasks, respectively.
Since the consensus reference alignments only allow us to assess how systems perform in comparison with one another, the proposed ranking is only a reference. Note that some of the correspondences in the consensus alignment may be erroneous (false positives) because all systems that agreed on it could be wrong (e.g., in erroneous correspondences with equivalent labels, which are not that uncommon in biomedical tasks). In addition, the consensus alignments will not be complete, because there are likely to be correct correspondences that no system is able to find, and there are a number of correspondences found by only one system (and therefore not in the consensus alignments) which may be correct. Nevertheless, the results with respect to the consensus alignments do provide some insights into the performance of the systems.
Overall, LogMap and LogMapBio are the systems that provide the closest set of correspondences to the consensus (not necessarily the best system) in both tasks. LogMap has a small set of unique correspondences as most of its correspondences are also suggested by its variant LogMapBio and vice versa. By contrast, AML and Wiktionary produce the highest number of unique correspondences in HP-MP and DOID-ORDO respectively, and the second-highest inversely. Nonetheless, Wiktionary suggests a very large number of correspondences with respect to the other systems which suggest that it may also include many subsumption and related correspondences and not only equivalence. All systems produce coherent alignments except for Wiktionary in the DOIDORDO task. In the OAEI 2019 Large Biomedical Ontologies track, 10 systems were able to complete at least one of the tasks within a 6 hours timeout. Eight systems were able to complete all six tasks.16 The evaluation results for the largest matching tasks are shown in Table 12.
The top-ranked systems by F-measure were respectively: AML and LogMap in Task 2; LogMap and LogMapBio in Task 4; and AML and LogMapBio in Task 6. 16 Check out the supporting scripts to reproduce the evaluation: https://github.com/ ernestojimenezruiz/oaei-evaluation
Interestingly, the use of background knowledge led to an improvement in recall from LogMapBio over LogMap in all tasks, but this came at the cost of precision, resulting in the two variants of the system having very similar F-measures.
The effectiveness of all systems decreased from small fragments to whole ontologies tasks.17 One reason for this is that with larger ontologies there are more plausible correspondence candidates, and thus it is harder to attain both a high precision and a high recall. In fact, this same pattern is observed moving from the FMA-NCI to the FMA-SNOMED to the SNOMED-NCI problem, as the size of the task also increases. Another reason is that the very scale of the problem constrains the matching strategies that systems can employ: AML for example, forgoes its matching algorithms that are computationally more complex when handling very large ontologies, due to efficiency concerns. 17 http://www.cs.ox.ac.uk/isg/projects/SEALS/oaei/2019/results/
The size of the whole ontologies tasks proved a problem for a some of the systems, which were unable to complete them within the allotted time: POMAP++ and SANOM.
With respect to alignment coherence, as in previous OAEI editions, only two distinct systems have shown alignment repair facilities: AML, LogMap and its LogMapBio variant. Note that only LogMap and LogMapBio are able to reduce to a minimum the number of unsatisfiable classes across all tasks, missing 3 unsatisfiable classes in the worst case (whole FMA-NCI task). For the AGM correspondences the ELK reasoner could not complete the classification over the integrated ontology within the allocated time.
As the results tables show, even the most precise alignment sets may lead to a huge number of unsatisfiable classes. This proves the importance of using techniques to assess the coherence of the generated alignments if they are to be used in tasks involving reasoning. We encourage ontology matching system developers to develop their own repair techniques or to use state-of-the-art techniques such as Alcomo [39], the repair module of LogMap (LogMap-Repair) [31] or the repair module of AML [43], which have worked well in practice [33, 24]. 4.7
This year, 5 systems registered to participate in the MultiFarm track: AML, EVOCROS, Lily, LogMap and Wiktionary. This number slightly decreases with respect to the last campaign (6 in 2018, 8 in 2017, 7 in 2016, 5 in 2015, 3 in 2014, 7 in 2013, and 7 in 2012). The reader can refer to the OAEI papers for a detailed description of the strategies adopted by each system. In fact, most systems still adopt a translation step before the matching itself. However, a few systems had issues when evaluated: i) EVOCROS encountered problems to complete a single matching task; and ii) Lily has generated mostly empty alignments.
The Multifarm evaluation results based on the blind dataset are presented in Table 13. They have been computed using the Alignment API 4.9 and can slightly differ from those computed with the SEALS client. We haven’t applied any threshold on the results. We do not report the results of non-specific systems here, as we could observe in the last campaigns that they can have intermediate results in the “same ontologies” task (ii) and poor performance in the “different ontologies” task (i).
AML outperforms all other systems in terms of F-measure for task i) (same behaviour than last year). In terms of precision, the systems have relatively similar results. With respect to the task ii) LogMap has the best performance. AML and LogMap have participated last year. Comparing the results from last year, in terms F-measure (cases of type i), AML maintains its overall performance (.45 in 2019, .46 in 2018, .46 in 2017, .45 in 2016 and .47 in 2015). The same could be observed for LogMap (.37 in 2018, .36 in 2017, and .37 in 2016).
In terms of performance, the F-measure for blind tests remains relatively stable across campaigns. AML and LogMap keep their positions and have similar F-measure with respect to the previous campaigns. As observed in previous campaigns, systems privilege precision over recall, and the results are expectedly below the ones obtained for the original Conference dataset. Cross-lingual approaches remain mainly based on translation strategies and the combination of other resources (like cross-lingual links
Time #pairs SizTeypePr(ei)c.– 22 Fte-smts. per pRaeicr. SizTeypeP(rieic).– 2 tFes-tms .per pRaierc. 236 55 8.18 .72 (.72) .45 (.45) .34 (.34) 33.40 .93 (.95) .27 (.28) .17 (.16) 49 55 6.99 .72 (.72) .37 (.37) .25 (.25) 46.80 .95 (.96) .41 (.42) .28 (.28) 785 23 4.91 .76 (.79) .31 (.33) .21 (.22) 9.24 .94 (.96) .12 (.12) .07 (.06) in Wikipedia, BabelNet, etc.) while strategies such as machine learning, or indirect alignment composition remain under-exploited. This year the Link Discovery track counted one participant in the Linking test case (AML) and three participants in the Spatial test case: AML, Silk and RADON. Those were the exact same systems (and versions) that participated on OAEI 2018.
In the Linking test case, AML perfectly captures all the correct links while not producing wrong ones, thus obtaining perfect precision and a recall (1.0) in both the Sandbox and Mainbox datasets. It required 9.7s and 360s, respectively, to complete the two tasks. The results can also be found in HOBBIT platform (https://tinyurl. com/yywwlsmt - Login as Guest).
We divided the Spatial test cases into four suites. In the first two suites (SLL and LLL), the systems were asked to match LineStrings to LineStrings considering a given relation for 200 and 2K instances for the TomTom and Spaten datasets. In the last two tasks (SLP, LLP), the systems were asked to match LineStrings to Polygons (or Polygons to LineStrings depending on the relation) again for both datasets. Since the precision, recall and f-measure results from all systems were equal to 1.0, we are only presenting results regarding the time performance. The time performance of the matching systems in the SLL, LLL, SLP and LLP suites are shown in Figures 2-3. The results can also be found in HOBBIT platform (https://tinyurl.com/y4vk6htq Login as Guest).
In the SLL suite, RADON has the best performance in most cases except for the Touches and Intersects relations, followed by AML. Silk seems to need the most time, particularly for Touches and Intersects relations in the TomTom dataset and Overlaps in both datasets.
In the LLL suite we have a more clear view of the capabilities of the systems with the increase in the number of instances. In this case, RADON and Silk have similar behavior as in the the small dataset, but it is more clear that the systems need much more time to match instances from the TomTom dataset. RADON has still the best performance in most cases. AML has the next best performance and is able to handle some cases better than other systems (e.g. Touches and Intersects), however, it also hits the platform time limit in the case of Disjoint.
In the SLP suite, in contrast to the first two suites, RADON has the best performance for all relations. AML and Silk have minor time differences and, depending on the case, one is slightly better than the other. All the systems need more time for the TomTom dataset but due to the small size of the instances the time difference is minor.
In the LLP suite, RADON again has the best performance in all cases. AML hits the platform time limit in Disjoint relations on both datasets and is better than Silk in most cases except Contains and Within on the TomTom dataset where it needs an excessive amount of time.
Taking into account the executed test cases we can identify the capabilities of the tested systems as well as suggest some improvements. All the systems participated in most of the test cases, with the exception of Silk which did not participate in the Covers and Covered By test cases.
RADON was the only system that successfully addressed all the tasks, and had the best performance for the SLP and LLP suites, but it can be improved for the Touches and Intersects relations for the SLL and LLL suites. AML performs extremely well in most cases, but can be improved in the cases of Covers/Covered By and Contains/Within when it comes to LineStrings/Polygons Tasks and especially in Disjoint relations where it hits the platform time limit. Silk can be improved for the Touches, Intersects and Overlaps relations and for the SLL and LLL tasks and for the Disjoint relation in SLP and LLP Tasks.
In general, all systems needed more time to match the TomTom dataset than the Spaten one, due to the smaller number of points per instance in the latter. Comparing the LineString/LineString to the LineString/Polygon Tasks we can say that all the systems needed less time for the first for the Contains, Within, Covers and Covered by relations, more time for the Touches, Instersects and Crosses relations, and approximately the same time for the Disjoint relation. This year, the SPIMBENCH track counted four participants: AML, Lily, LogMap and FTRLIM. FTRLIM participated for the first time this year while AML, Lily, and LogMap also participated the previous years. The evaluation results of the track are shown in Table 14. The results can also be found in HOBBIT platform (https: //tinyurl.com/yxhsw48c - Login as Guest).
System Precision Recall F-measure Time (ms)
Lily and FTRLIM had the best performance overall both in terms of F-measure and run time. Notably, their run time scaled very well with the increase in the number of instances. Lily, FTRLIM, and AML had a higher recall than precision, while Lily and FTRLIM had a full recall. By contrast, LogMap had the highest precision but lowest recall of all the systems. AML and LogMap had a similar run time for the Sandbox task, but the latter scaled better with the increase in the number of instances. We evaluated all SEALS participants in the OAEI (even those not registered for the track) on a very small matching task18. This revealed that not all systems were able to handle the task, and in the end, only the following systems were evaluated: AGM, AML, DOME, FCAMap-KG, LogMap, LogMapBio, LogMapKG, LogMapLt, POMap++, Wiktionary. Out of those only LogMapBio, LogMapLt and POMap++ were not registered for this track. In comparison to last year, more matchers participate and return meaningful correspondences. Moreover there are systems which especially focus on the knowledge graph track e.g. FCAMap-KG and LogMapKG.
Table 15 shows the aggregated results for all systems, including the number of tasks in which they were able to generate a non-empty alignment (#tasks) and the average number of generated correspondences in those tasks (size). In addition to the global average precision, F-measure, and recall results, in which tasks where systems produced empty alignments were counted, we also computed F-measure and recall ignoring empty alignments which are shown between parentheses in the table, where applicable.
Nearly all systems were able to generate class correspondences. In terms of Fmeasure, AML is the best one (when considering only completed test cases). Many matchers were also able to beat the baseline. The highest recall is about 0.77 which shows that some class correspondences are not easy to find.
In comparison to last year, more matchers are able to produce property correspondences. Only the systems of the LogMap family and POMAP++ do not return any alignments. While Wiktionary and FCAMap-KG achieve an F-Measure of 0.98, other systems need more improvement here because they are not capable of beating the baseline (mostly due to low recall).
With respect to instance correspondences, AML and DOME are the best systems, but they outperform the baselines only by a small margin. On average, the systems returned between 3,000 and 8,000 instance alignments. Only LogMapKG returned nearly 30,000 mappings. This is interesting because it should be focused on generating only 1:1 alignments, but deviates here.
We also analyzed the arity of the resulting alignments because in the knowledge graph track it is probably better to focus on a 1:1 mapping. Such a strict mapping is returned by the following systems: AGM, baselineLabel, DOME and POMAP++. LogMap and LogMapBio return a few correspondences with same source or target in only two test cases. BaselineAltLabel, FCAMap-KG and Wiktionary returned some n:m mappings in all test cases. AML and LogMapLt returned more of those and LogMapKG has the highest amount of n:m mappings.
When analyzing the confidence values of the alignments, it turns out that most matchers set it to 1 (AGM,baselineAltLabel, baselineLabel, FCAMap-KG, LogMapLt, 18 http://oaei.ontologymatching.org/2019/results/knowledgegraph/ small_test.zip
AGM 10:47:38 AML 0:45:46 baselineAltLabel 0:11:48 baselineLabel 0:12:30 DOME 1:05:26 FCAMap-KG 1:14:49 LogMap 0:15:43 LogMapBio 2:31:01 LogMapKG 2:26:14 LogMapLt 0:07:28 POMAP++ 0:14:39 Wiktionary 0:20:14 AGM 10:47:38 AML 0:45:46 baselineAltLabel 0:11:48 baselineLabel 0:12:30 DOME 1:05:26 FCAMap-KG 1:14:49 LogMap 0:15:43 LogMapBio 2:31:01 LogMapKG 2:26:14 LogMapLt 0:07:28 POMAP++ 0:14:39 Wiktionary 0:20:14 AGM 10:47:38 AML 0:45:46 baselineAltLabel 0:11:48 baselineLabel 0:12:30 DOME 1:05:26 FCAMap-KG 1:14:49 LogMap 0:15:43 LogMapBio 2:31:01 LogMapKG 2:26:14 LogMapLt 0:07:28 POMAP++ 0:14:39 Wiktionary 0:20:14 AGM 10:47:38 AML 0:45:46 baselineAltLabel 0:11:48 baselineLabel 0:12:30 DOME 1:05:26 FCAMap-KG 1:14:49 LogMap 0:15:43 LogMapBio 2:31:01 LogMapKG 2:26:14 LogMapLt 0:07:28 POMAP++ 0:14:39 Wiktionary 0:20:14
Class performance 5 14.6 0.23 0.09 0.06) 4 27.5 0.78 (0.98) 0.69 (0.86) 0.61 (0.77) 5 16.4 1.0 0.74 0.59 5 16.4 1.0 0.74 0.59 4 22.5 0.74 (0.92) 0.62 (0.77) 0.53 (0.66) 5 18.6 1.0 0.82 0.70 5 26.0 0.95 0.84 0.76) 5 26.0 0.95 0.84 0.76) 5 26.0 0.95 0.84 0.76) 4 23.0 0.80 (1.0) 0.56 (0.70) 0.43 (0.54) 5 2.0 0.0 0.0 0.0 5 21.4 1.0 0.8 0.67 Wiktionary). AML and LogMapKG set it higher than 0.6 whereas only DOME uses the full range between zero and one. LogMap and LogMapBio uses a range of 0.3 and 0.8. The confidences were analyzed with the MELT dashboard19 [28].
Regarding runtime, AGM (10:47:38) was the slowest system, followed by LogMapKG and LogMapBio which were much faster. Besides AGM all five test cases could be completed in under 3 hours.
This year, three systems participated in the Interactive matching track. They are ALIN, AML, and LogMap. Their results are shown in Table 16 and Figure 4 for both Anatomy and Conference datasets.
The table includes the following information (column names within parentheses): – The performance of the system: Precision (Prec.), Recall (Rec.) and F-measure (Fm.) with respect to the fixed reference alignment, as well as Recall+ (Rec.+) for the Anatomy task. To facilitate the assessment of the impact of user interactions, we also provide the performance results from the original tracks, without interaction (line with Error NI). – To ascertain the impact of the oracle errors, we provide the performance of the system with respect to the oracle (i.e., the reference alignment as modified by the errors introduced by the oracle: Precision oracle (Prec. oracle), Recall oracle (Rec. oracle) and F-measure oracle (F-m. oracle). For a perfect oracle these values match the actual performance of the system. – Total requests (Tot Reqs.) represents the number of distinct user interactions with the tool, where each interaction can contain one to three conflicting correspondences, that could be analysed simultaneously by a user. – Distinct correspondences (Dist. Mapps) counts the total number of correspondences for which the oracle gave feedback to the user (regardless of whether they were submitted simultaneously, or separately). – Finally, the performance of the oracle itself with respect to the errors it introduced can be gauged through the positive precision (Pos. Prec.) and negative precision (Neg. Prec.), which measure respectively the fraction of positive and negative answers given by the oracle that are correct. For a perfect oracle these values are equal to 1 (or 0, if no questions were asked).
The figure shows the time intervals between the questions to the user/oracle for the different systems and error rates. Different runs are depicted with different colors.
The matching systems that participated in this track employ different userinteraction strategies. While LogMap, and AML make use of user interactions exclusively in the post-matching steps to filter their candidate correspondences, ALIN can also add new candidate correspondences to its initial set. LogMap and AML both request feedback on only selected correspondences candidates (based on their similarity 19 http://oaei.ontologymatching.org/2019/results/knowledgegraph/ knowledge_graph_dashboard.html 0.871 0.443 0.587 0.914 0.695 0.79 0.809 0.658 0.725 0.715 0.631 0.67 0.636 0.605 0.62 0.841 0.659 0.739 0.91 0.698 0.79 0.846 0.687 0.758 0.783 0.67 0.721 0.721 0.646 0.681 – – – – – – – – – – – – – – NI stands for non-interactive, and refers to the results obtained by the matching system in the original track. patterns or their involvement in unsatisfiabilities) and AML presents one correspondence at a time to the user. ALIN and LogMap can both ask the oracle to analyze several conflicting correspondences simultaneously.
The performance of the systems usually improves when interacting with a perfect oracle in comparison with no interaction. ALIN is the system that improves the most, because its high number of oracle requests and its non-interactive performance was the lowest of the interactive systems, and thus the easiest to improve.
Although system performance deteriorates when the error rate increases, there are still benefits from the user interaction—some of the systems’ measures stay above their non-interactive values even for the larger error rates. Naturally, the more a system relies on the oracle, the more its performance tends to be affected by the oracle’s errors.
The impact of the oracle’s errors is linear for ALIN, and AML in most tasks, as the F-measure according to the oracle remains approximately constant across all error rates. It is supra-linear for LogMap in all datasets.
Another aspect that was assessed, was the response time of systems, i.e., the time between requests. Two models for system response times are frequently used in the literature [11]: Shneiderman and Seow take different approaches to categorize the response times taking a task-centered view and a user-centered view respectively. According to task complexity, Shneiderman defines response time in four categories: typing, mouse movement (50-150 ms), simple frequent tasks (1 s), common tasks (2-4 s) and complex tasks (8-12 s). While Seow’s definition of response time is based on the user expectations towards the execution of a task: instantaneous (100-200 ms), immediate (0.5-1 s), continuous (2-5 s), captive (7-10 s). Ontology alignment is a cognitively demanding task and can fall into the third or fourth categories in both models. In this regard the response times (request intervals as we call them above) observed in all datasets fall into the tolerable and acceptable response times, and even into the first categories, in both models. The request intervals for AML, LogMap and XMAP stay at a few milliseconds for most datasets. ALIN’s request intervals are higher, but still in the tenth of second range. It could be the case, however, that a user would not be able to take advantage of these low response times because the task complexity may result in higher user response time (i.e., the time the user needs to respond to the system after the system is ready). Three systems were able to generate complex correspondences: AMLC, AROA [53], and CANARD. The results for the other systems are reported in terms of simple alignments. The results of the systems on the five test cases are summarized in Table 17.
With respect to the Hydrography test case, only AMLC can generate two correct complex correspondences which are stating that a class in the source ontology is equivalent to the union of two classes in the target ontology. Most of the systems achieved fair results in terms of precision, but the low recall reflects that the current ontology alignment systems still need to be improved to find more complex relations.
In terms of GeoLink test cases, the real-world instance data from GeoLink Project is also populated into the ontology in order to enable the systems that depend on instancebased matching algorithms to evaluate their performance. There are three alignment systems that generate complex alignments in GeoLink Benchmark, which are AMLC, AROA, and CANARD. AMLC didn’t find any correct complex alignment, while AROA and CARARD achieved relatively good performance. One of the reasons may be that these two systems are instance-based systems, which rely on the shared instances between ontologies. In other words, the shared instance data between two ontologies would be helpful to the matching process.
In the Taxon test cases, only the output of LogMap, LogMapLt and CANARD could be used to rewrite source queries.
With respect to the Conference test cases although the performance in terms of precision and recall decreased for AMLC, AMLC managed to find more true positives than the last year. Since AMLC provides confidence, it could be possible to include confidence into the evaluation and this could improve the performance results. AMLC discovered one more kind of complex mappings: the union of classes.
A more detailed discussion of the results of each task can be found in the OAEI page for this track. For a second edition of complex matching in an OAEI campaign, and given the inherent difficulty of the task, the results and participation are promising albeit still modest. 5
In 2019, we witnessed a slight decrease in the number of participants in comparison with previous years, but with a healthy mix of new and returning systems. However, like last year, the distribution of participants by tracks was uneven.
The schema matching tracks saw abundant participation, but, as has been the trend of the recent years, little substantial progress in terms of quality of the results or run time of top matching systems, judging from the long-standing tracks. On the one hand, this may be a sign of a performance plateau being reached by existing strategies and algorithms, which would suggest that new technology is needed to obtain significant improvements. On the other hand, it is also true that established matching systems tend to focus more on new tracks and datasets than on improving their performance in longstanding tracks, whereas new systems typically struggle to compete with established ones.
The number of matching systems capable of handling very large ontologies has increased slightly over the last years, but is still relatively modest, judging from the Large Biomedical Ontologies track. We will aim at facilitating participation in future editions of this track by providing techniques to divide the matching tasks in manageable subtasks (e.g., [30]).
According to the Conference track there is still need for an improvement with regard to the ability of matching systems to match properties. To assist system developers in tackling this aspect we provided a more detailed evaluation in terms of the analysis of the false positives per matching system (available on the Conference track web page). However, this could be extended by the inspection of the reasons why the matching system found the given false positives.As already pointed out last year, less encouraging is the low number of systems concerned with the logical coherence of the alignments they produce, an aspect which is critical for several semantic web applications. Perhaps a more direct approach is needed to promote this topic, such as providing a more indepth analysis of the causes of incoherence in the evaluation or even organizing a future track focusing on logical coherence alone.
The consensus-based evaluation in the Disease and Phenotype track offers limited insights into performance, as several matching systems produce a number of unique correspondences which may or may not be correct. In the absence of a true reference alignment, future evaluation should seek to determine whether the unique correspondences contain indicators of correctness, such as semantic similarity, or appear to be noise.
Despite the quite promising results obtained by matching systems for the Biodiversity and Ecology track, the most important observation is that none of the systems has been able to detect mappings established by the experts. Detecting such correspondences requires the use of domain-specific core knowledge that captures biodiversity concepts. We expect this domain-specific background to be integrated in future versions of the systems.
The interactive matching track also witnessed a small number of participants. Three systems participated this year. This is puzzling considering that this track is based on the Anatomy and Conference test cases, and those tracks had 13 participants. The process of programmatically querying the Oracle class used to simulate user interactions is simple enough that it should not be a deterrent for participation, but perhaps we should look at facilitating the process further in future OAEI editions by providing implementation examples.
The complex matching track opens new perspectives in the field of ontology matching. Tackling complex matching automatically is extremely challenging, likely requiring profound adaptations from matching systems, so the fact that there were three participants that were able to generate complex correspondences in this track should be seen as a positive sign of progress to the state of the art in ontology matching. This year automatic evaluation has been introduced following an instance-based comparison approach.
The instance matching tracks and the new instance and schema matching track counted few participants, as has been the trend in recent years. Part of the reason for this is that several of these tracks ran on the HOBBIT platform, and the transition from SEALS to HOBBIT has not been as easy as we might desire. Thus, participation should increase next year as systems become more familiar with the HOBBIT platform and have more time to do the migration. Furthermore, from an infrastructure point of view, the HOBBIT SDK will make the developing and debugging phase easier, and the Maven-based framework will facilitate submission. However, another factor behind the reduced participation in the instance matching tracks lies with their specialization. New schema matching tracks such as Biodiversity and Ecology typically demand very little from systems that are already able to tackle long-standing tracks such as Anatomy, whereas instance matching tracks such as Link Discovery and last year’s Process Model Matching, are so different from one another that each requires dedicated development time to tackle. Thus, in future OAEI editions we should consider publishing new instance matching (and other more specialized) datasets with more time in advance, to give system developers adequate time to tackle them. Equally critical will be to ensure stability by maintaining instance matching tracks and datasets over multiple OAEI editions, so that participants can build upon the development of previous years.
Automatic instance-matching benchmark generation algorithms have been gaining popularity, as evidenced by the fact that they are used in all three instance matching tracks of this OAEI edition. One aspect that has not been addressed in such algorithms is that, if the transformation is too extreme, the correspondence may be unrealistic and impossible to detect even by humans. As such, we argue that human-in-the-loop techniques can be exploited to do a preventive quality-checking of generated correspondences, and refine the set of correspondences included in the final reference alignment.
In the knowledge graph track, we could observe that simple baselines are still hard to beat – which was also the case in other tracks when they were still new. We expect more sophisticated and powerful implementations in the next editions.
Like in previous OAEI editions, most participants provided a description of their systems and their experience in the evaluation, in the form of OAEI system papers. These papers, like the present one, have not been peer reviewed. However, they are full contributions to this evaluation exercise, reflecting the effort and insight of matching systems developers, and providing details about those systems and the algorithms they implement.
The Ontology Alignment Evaluation Initiative will strive to remain a reference to the ontology matching community by improving both the test cases and the testing methodology to better reflect actual needs, as well as to promote progress in this field. More information can be found at: http://oaei.ontologymatching.org.
We warmly thank the participants of this campaign. We know that they have worked hard to have their matching tools executable in time and they provided useful reports on their experience. The best way to learn about the results remains to read the papers that follow.
We are grateful to the Universidad Polite´cnica de Madrid (UPM), especially to Nandana Mihindukulasooriya and Asuncio´n Go´mez Pe´rez, for moving, setting up and providing the necessary infrastructure to run the SEALS repositories.
We are also grateful to Martin Ringwald and Terry Hayamizu for providing the reference alignment for the anatomy ontologies and thank Elena Beisswanger for her thorough support on improving the quality of the dataset.
We thank Andrea Turbati and the AGROVOC team for their very appreciated help with the preparation of the AGROVOC subset ontology. We are also grateful to Catherine Roussey and Nathalie Hernandez for their help on the Taxon alignment.
We also thank for their support the past members of the Ontology Alignment Evaluation Initiative steering committee: Je´roˆme Euzenat (INRIA, FR), Yannis Kalfoglou (Ricoh laboratories, UK), Miklos Nagy (The Open University,UK), Natasha Noy (Google Inc., USA), Yuzhong Qu (Southeast University, CN), York Sure (Leibniz Gemeinschaft, DE), Jie Tang (Tsinghua University, CN), Heiner Stuckenschmidt (Mannheim Universita¨t, DE), George Vouros (University of the Aegean, GR).
Ca´ssia Trojahn dos Santos has been partially supported by the CNRS Blanc project RegleX-LD.
Daniel Faria was supported by the EC H2020 grant 676559 ELIXIREXCELERATE and the Portuguese FCT Grant 22231 BioData.pt, co-financed by FEDER.
Ernesto Jimenez-Ruiz has been partially supported by the SIRIUS Centre for Scalable Data Access (Research Council of Norway, project no.: 237889) and the AIDA project (Alan Turing Institute).
Catia Pesquita was supported by the FCT through the LASIGE Strategic Project (UID/CEC/00408/2013) and the research grant PTDC/EEI-ESS/4633/2014.
Irini Fundulaki and Tzanina Saveta were supported by the EU’s Horizon 2020 research and innovation programme under grant agreement No 688227 (Hobbit).
Jana Vatasˇcˇinova´ and Ondrˇej Zamazal were supported by the CSF grant no. 1823964S.
Patrick Lambrix and Huanyu Li have been supported by the Swedish e-Science Research Centre (SeRC), the Swedish Research Council (Vetenskapsra˚det, dnr 201804147) and the Swedish National Graduate School in Computer Science (CUGS).
The Biodiversity and Ecology track has been partially funded by the German Research Foundation in the context of the GFBio Project (grant No. SE 553/7-1) and the CRC 1076 AquaDiva, the Leitprojekt der Fraunhofer Gesellschaft in the context of the MED2ICIN project (grant No. 600628) and the German Network for Bioinformatics Infrastructure - de.NBI (grant No. 031A539B). 2017. In Proceedings of the 12th International Workshop on Ontology Matching, Vienna, Austria, pages 61–113, 2017. 2. Manel Achichi, Michelle Cheatham, Zlatan Dragisic, Jerome Euzenat, Daniel Faria, Alfio Ferrara, Giorgos Flouris, Irini Fundulaki, Ian Harrow, Valentina Ivanova, Ernesto Jime´nezRuiz, Elena Kuss, Patrick Lambrix, Henrik Leopold, Huanyu Li, Christian Meilicke, Stefano Montanelli, Catia Pesquita, Tzanina Saveta, Pavel Shvaiko, Andrea Splendiani, Heiner Stuckenschmidt, Konstantin Todorov, Ca´ssia Trojahn, and Ondrej Zamazal. Results of the ontology alignment evaluation initiative 2016. In Proceedings of the 11th International Ontology matching workshop, Kobe (JP), pages 73–129, 2016. 3. Jose´ Luis Aguirre, Bernardo Cuenca Grau, Kai Eckert, Je´roˆ me Euzenat, Alfio Ferrara, Robert Willem van Hague, Laura Hollink, Ernesto Jime´nez-Ruiz, Christian Meilicke, Andriy Nikolov, Dominique Ritze, Franc¸ois Scharffe, Pavel Shvaiko, Ondrej Sva´b-Zamazal, Ca´ssia Trojahn, and Benjamin Zapilko. Results of the ontology alignment evaluation initiative 2012. In Proceedings of the 7th International Ontology matching workshop, Boston (MA, US), pages 73–115, 2012. 4. Alsayed Algergawy, Michelle Cheatham, Daniel Faria, Alfio Ferrara, Irini Fundulaki, Ian Harrow, Sven Hertling, Ernesto Jime´nez-Ruiz, Naouel Karam, Abderrahmane Khiat, Patrick Lambrix, Huanyu Li, Stefano Montanelli, Heiko Paulheim, Catia Pesquita, Tzanina Saveta, Daniela Schmidt, Pavel Shvaiko, Andrea Splendiani, E´ lodie Thie´blin, Ca´ssia Trojahn, Jana Vatascinova´, Ondrej Zamazal, and Lu Zhou. Results of the ontology alignment evaluation initiative 2018. In Proceedings of the 13th International Workshop on Ontology Matching, Monterey (CA, US), pages 76–116, 2018. 5. Benhamin Ashpole, Marc Ehrig, Je´r oˆme Euzenat, and Heiner Stuckenschmidt, editors. Proc.
K-Cap Workshop on Integrating Ontologies, Banff (Canada), 2005. 6. Olivier Bodenreider. The unified medical language system (UMLS): integrating biomedical terminology. Nucleic Acids Research, 32:267–270, 2004. 7. Caterina Caracciolo, Je´roˆ me Euzenat, Laura Hollink, Ryutaro Ichise, Antoine Isaac, Ve´ronique Malaise´, Christian Meilicke, Juan Pane, Pavel Shvaiko, Heiner Stuckenschmidt, Ondrej Sva´b-Zamazal, and Vojtech Sva´tek. Results of the ontology alignment evaluation initiative 2008. In Proceedings of the 3rd Ontology matching workshop, Karlsruhe (DE), pages 73–120, 2008. 8. Michelle Cheatham, Zlatan Dragisic, Je´roˆ me Euzenat, Daniel Faria, Alfio Ferrara, Giorgos Flouris, Irini Fundulaki, Roger Granada, Valentina Ivanova, Ernesto Jime´nez-Ruiz, Patrick Lambrix, Stefano Montanelli, Catia Pesquita, Tzanina Saveta, Pavel Shvaiko, Alessandro Solimando, Ca´ssia Trojahn, and Ondrˇej Zamazal. Results of the ontology alignment evaluation initiative 2015. In Proceedings of the 10th International Ontology matching workshop, Bethlehem (PA, US), pages 60–115, 2015. 9. Michelle Cheatham, Dalia Varanka, Fatima Arauz, and Lu Zhou. Alignment of surface water ontologies: a comparison of manual and automated approaches. Journal of Geographical Systems, pages 1–23, 2019. 10. Bernardo Cuenca Grau, Zlatan Dragisic, Kai Eckert, Je´roˆ me Euzenat, Alfio Ferrara, Roger Granada, Valentina Ivanova, Ernesto Jime´nez-Ruiz, Andreas Oskar Kempf, Patrick Lambrix, Andriy Nikolov, Heiko Paulheim, Dominique Ritze, Franc¸ois Scharffe, Pavel Shvaiko, Ca´ssia Trojahn dos Santos, and Ondrej Zamazal. Results of the ontology alignment evaluation initiative 2013. In Pavel Shvaiko, Je´roˆ me Euzenat, Kavitha Srinivas, Ming Mao, and Ernesto Jime´nez-Ruiz, editors, Proceedings of the 8th International Ontology matching workshop, Sydney (NSW, AU), pages 61–100, 2013. 11. Jim Dabrowski and Ethan V. Munson. 40 years of searching for the best computer system response time. Interacting with Computers, 23(5):555–564, 2011. 41. Boris Motik, Rob Shearer, and Ian Horrocks. Hypertableau reasoning for description logics.
Journal of Artificial Intelligence Research, 36:165–228, 2009. 42. Heiko Paulheim, Sven Hertling, and Dominique Ritze. Towards evaluating interactive ontology matching tools. In Proceedings of the 10th Extended Semantic Web Conference, Montpellier (FR), pages 31–45, 2013. 43. Emanuel Santos, Daniel Faria, Catia Pesquita, and Francisco M Couto. Ontology alignment repair through modularization and confidence-based heuristics. PLoS ONE, 10(12):e0144807, 2015. 44. Tzanina Saveta, Evangelia Daskalaki, Giorgos Flouris, Irini Fundulaki, Melanie Herschel, and Axel-Cyrille Ngonga Ngomo. Pushing the limits of instance matching systems: A semantics-aware benchmark for linked data. In Proceedings of the 24th International Conference on World Wide Web, pages 105–106, New York, NY, USA, 2015. ACM. 45. Alessandro Solimando, Ernesto Jime´nez-Ruiz, and Giovanna Guerrini. Detecting and correcting conservativity principle violations in ontology-to-ontology mappings. In Proceedings of the International Semantic Web Conference, pages 1–16. Springer, 2014. 46. Alessandro Solimando, Ernesto Jimenez-Ruiz, and Giovanna Guerrini. Minimizing conservativity violations in ontology alignments: Algorithms and evaluation. Knowledge and Information Systems, 2016. 47. Christian Strobl. Encyclopedia of GIS, chapter Dimensionally Extended Nine-Intersection
Model (DE-9IM), pages 240–245. Springer, 2008. 48. York Sure, Oscar Corcho, Je´roˆ me Euzenat, and Todd Hughes, editors. Proceedings of the
Workshop on Evaluation of Ontology-based Tools (EON), Hiroshima (JP), 2004. 49. E´lodie Thie´blin. Do competency questions for alignment help fostering complex correspondences? In Proceedings of the EKAW Doctoral Consortium 2018, 2018. 50. E´lodie Thie´blin, Fabien Amarger, Ollivier Haemmerle´, Nathalie Hernandez, and Ca´ssia Trojahn dos Santos. Rewriting SELECT SPARQL queries from 1: n complex correspondences. In Proceedings of the 11th International Workshop on Ontology Matching, pages 49–60, 2016. 51. Elodie Thie´blin, Michelle Cheatham, Cassia Trojahn, Ondrej Zamazal, and Lu Zhou. The First Version of the OAEI Complex Alignment Benchmark. In Proceedings of the International Semantic Web Conference (Posters and Demos), 2018. 52. Ondrˇej Zamazal and Vojteˇch Sva´tek. The ten-year ontofarm and its fertilization within the onto-sphere. Web Semantics: Science, Services and Agents on the World Wide Web, 43:46– 53, 2017. 53. Lu Zhou, Michelle Cheatham, and Pascal Hitzler. Towards association rule-based complex ontology alignment. In Proceedings of the 9th Joint International Semantic Technology Conference JIST 2019, Hangzhou, China, November 25, in press, 2019. 54. Lu Zhou, Michelle Cheatham, Adila Krisnadhi, and Pascal Hitzler. A complex alignment benchmark: Geolink dataset. In Proceedings of the 17th International Semantic Web Conference, Monterey (CA, USA), pages 273–288, 2018. 55. Lu Zhou, Michelle Cheatham, Adila Krisnadhi, and Pascal Hitzler. Geolink dataset: a complex alignment benchmark from real-world ontology. Data Intelligence, in press, 2019.