This paper presents the results of the ALOD2vec Matcher in the Ontology Alignment Evaluation Initiative (OAEI) 2021. The matching system exploits a Web-scale dataset, i.e. WebIsALOD, as background knowledge source. In order to make use of the dataset, the RDF2vec approach is applied to derive embeddings for each concept available in the dataset. ALOD2vec Matcher participated in the OAEI 2018 and 2020 campaigns before. This is the system's third participation.3 The ALOD2vec Matcher is an element-level, label-based matcher which uses a large-scale Web-crawled RDF dataset of hypernymy relations as general purpose background knowledge. The dataset contains many tail-entities as well as instance data such as persons or places which cannot be found in common thesauri. In order to exploit the external dataset, a neural language model approach is used to obtain a vector for each concept contained in the dataset. This matching system system was initially introduced at the OAEI 2018 [13] and also participated in the 2020 campaign [10]. The implementation is based on the Matching EvaLuation Toolkit [6] as well as the KGvec2go [11] REST API to obtain vector representations via a Web API.
1.1
1.2
After the basic concepts of this matcher are introduced (Foundations ), the speci c techniques applied are presented.
3 Copyright © 2021 for this paper by its authors. Use permitted under Creative
Commons License Attribution 4.0 International (CC BY 4.0).
WebIsALOD Dataset A frequent problem that occurs when working with
external background knowledge is the fact that less common entities are not contained
within a knowledge base. The WebIsA [
RDF2vec The background dataset can be viewed as a very large knowledge
graph; in order to obtain a similarity score for nodes and edges in that graph, the
RDF2vec [
The rst step is to link the obtained labels from the ontology to concepts in
the WebIsALOD dataset. Therefore, string operations are performed on the label
4 see http://commoncrawl.org/
5 see http://webisa.webdatacommons.org/concept/european_union_
6 see http://webisa.webdatacommons.org/
7 For a good overview of the RDF2vec approach and its applications, refer to
http://www.rdf2vec.org/
8 see http://kgvec2go.org/api.html
and it is checked whether the label is available in WebIsALOD. If it cannot be
found, a token-lookup is performed. Given two entities e1 and e2, the matcher
uses their textual labels to link them to concepts e01 and e02 in the external
dataset. Afterwards, the embedding vectors ve01 and ve02 of the linked concepts
(e01 and e02) are retrieved via a Web request and the cosine similarity between
those is calculated. Hence: sim(e1; e2) = simcosine(ve01 ; ve02 ). If sim(e1; e2) > t
where t is a threshold in the range of 0 and 1, a correspondence is added to a
temporary alignment. In a last step, a one-to-one arity is enforced by applying
a Maximum Weight Bipartite [
In order to consume the vectors in Java, a client has been implemented and contributed to the MELT-ML module. The KGvec2go REST API can now be accessed though class KGvec2goClient. Even though this matcher only uses the WebIsALOD dataset, the implementation supports all datasets accessible on KGvec2go. The extension is available by default in MELT 2.6.
Instance Matching After classes and properties have been matched, instances are matched using a string index. The con dence score assigned to instances belonging to matched classes is higher than that of matches between instances belonging to non-matched classes.
Explainability ALOD2vec Matcher provides an explanation for every
correspondence that is added to the nal alignment. Therefore, the extension
capabilities of the alignment format [
For the 2021 campaign, the matching system was adapted to use the latest MELT release and was packaged as MELT Web Docker10 container. The 2021 implementation is publicly available on GitHub.11 2 2.1
On the anatomy dataset, the system scores a precision of 0.828, a recall of 0.766, and an F1 of 0.796. 2.2
On the conference track, the matcher achieves a recall of 0.49 and a precision of 0.64. The overall F1 score on ra1-M3 was 0.59. 2.3
Since the WebIsALOD dataset is only available in English, the focus of the ALOD2vec Matcher is on monolingual matching tasks. 2.4
In its current version, the LargeBio track is too large for the matching system's architecture. There is a tradeo in package size and runtime performance (a large package with all vectors matches faster than the submitted small package which obtains vectors at runtime from KGvec2go). The current architecture of ALOD2vec Matcher is not intended for large-scale matching { however, the matching algorithm itself could be used for large-scale matching. 2.5
The system could complete all matching tasks in time. As in the previous year,
this matcher obtains the second best results achieving almost the same score as
the Wiktionary Matcher 2021 [
This year, a new track was added to the OAEI: The Common Knowledge Graph
Track [
In this paper, we presented the newest version of the ALOD2vec Matcher, a matcher utilizing an RDF2vec vector representation of the WebIsALOD dataset, as well as its results in the 2021 OAEI. In the future, the matching system could be improved by using another, potentially larger or newer, hypernymy database, by exploiting other embedding algorithms, and by adding further matching strategies to the overall algorithms such as checking of logical constraints.