In this paper, we introduce the ALOD2Vec Matcher, an ontology matching tool that exploits a Web-scale data set, i.e., WebIsALOD, as external knowledge source. In order to make use of the data set, the RDF2Vec approach is chosen to derive embeddings for each concept available in the data set. We show that it is possible to use very large RDF graphs as external background knowledge source for the task of ontology matching.
1.1
The ALOD2Vec Matcher is an element-level, label-based matcher which uses a large-scale Web-crawled RDF data set of hypernymy relations as background knowledge. One advantage of that data set is the inclusion of many tail-entities, as well as instance data, such as persons or places, which cannot be found in thesauri. In order to make use of the external data set, a neural language model approach is used to calculate an embedding vector for each concept contained in it.
Given two entities e1 and e2, the matcher uses their textual labels to link them to concepts e01 and e02 in the external data set. Then, the pre-calculated embedding vectors ve01 and ve02 of the linked concepts (e01 and e02) are retrieved and the cosine similarity between those is calculated. Hence: sim(e1; e2) = simcosine(ve01 ; ve02 ). The resulting alignment is homogenous, i.e., classes, object properties, and datatype properties are handled separately. In addition, the matcher enforces a oneto-many matching restriction. 1.2
For the alignment process, the matcher retrieves textual descriptions of all
elements of the ontologies to be matched. A lter adds all simple string matches
to the nal alignment in order to increase the performance. The remaining
labels are linked to concepts of the background data set, are compared, and the
best solution is added to the nal alignment. A high-level view of the system is
depicted in gure 1.
? Supported by SAP SE.
WebIsALOD Data Set When working with knowledge bases in order to
exploit the contained knowledge in applications, a frequent problem is the fact
that less common entities are not contained within the knowledge base. The
WebIsA [
RDF2Vec The background data set can be viewed as a very large knowledge
graph; in order to obtain a similarity score for nodes in that graph, the RDF2Vec
[
Linking The rst step is to link the obtained labels from the ontology to concepts in the WebIsALOD data set. Therefore, string operations are performed on the label and it is checked whether the label is available in WebIsALOD. If it cannot be found, labels consisting of multiple words are truncated from the right, and the process is repeated to check for sub-concepts. For example, the label United Nations Peacekeeping Mission in Mali cannot be found in WebIsALOD. Therefore, it is truncated until the longest label from the left is found { in this case United Nations. The process is repeated until all tokens are processed. The resulting concepts for the given label are: United Nations4, peacekeeping mission5, and Mali 6.
Similarity Calculation As stated before, labels are linked to concepts, their vectors are retrieved, and the cosine similarity between them is used as similarity score.
There are cases in which parts of a label cannot be found, however, for example in tubule macula and in macula lutea both times only macula can be found using the WebIsALOD data set. If only the found concepts would be used to calculate the similarity between the concepts, a perfect score would be obtained because sim(macula; macula) = 1:0. This is not precise as the approach does not allow to discriminate between perfect matches due to incomplete linking and real perfect matches. Therefore, a penalty factor p 2 [0; 1] is introduced that is to be multiplied with the nal similarity score and which lowers the score for incomplete links; p = 0 indicates the maximal penalty, p = 1 indicates no penalty. The calculation of p is depicted in equation 1: p = 0:5 jF ound Concepts L1j + 0:5 jP ossible Concepts L1j jF ound Concepts L2j jP ossible Concepts L2j where L1 is the label of the rst concept and L2 is the label of the second one; jF ound Concepts Lij is the number of tokens for which a concept could be found (minus stopwords) and jP ossible Concepts Lij is the number of tokens of the label without stopwords. The penalty score is multiplied with the nal similarity score. Hence, incomplete linkages are penalized.
If two labels were matched to multiple concepts, a resolution is required. In this case the best average similarity is used: simaverage = ij2c1cj1 M axjjc22cj2 sim(c1i ; c2j ) jc1j where c1 and c2 represent two individual concepts and c1i , respectively c2j , represent the ith and jth sub-concept of c1 and c2; jc1j and jc2j are the number of sub-concepts of c1 and c2; c1 is the concept with more tokens. 4 see http://webisa.webdatacommons.org/concept/united_nations_ 5 see http://webisa.webdatacommons.org/concept/peacekeeping_mission_ 6 see http://webisa.webdatacommons.org/concept/_mali_ (1) (2)
Typically, there is more than one label to an entity of an ontology. Therefore, a score-matrix is used: Every label of an entity is linked and compared to every label of the other entity and the best score is returned.
RDF2Vec Con guration Parameters We generated 100 sentences of depth
8 for each node in the WebIsALOD data set for the training process of the model.
In order to have also sentences for nodes that do not have out-going edges, those
were identi ed and sentences were generated backwards and afterwards reversed.
The sentences were generated in a biased fashion [
For the Anatomy data set, the matcher achieves a higher recall and F1 score compared to the baseline solution. However, the true positives are mostly exact lexical matches or share many common tokens.
Concerning runtime-performance, ALOD2Vec Matcher performs in the upper half of all matchers that participated in the Anatomy track. 2.2
On the Conference data set, it can be seen that the matcher is better in aligning classes than in aligning properties. This is in line with the results reported for other matchers. In this case, it is due to fewer lexical matches in properties as well as the higher usage of non-nouns which cannot be properly linked to the background knowledge source. 2.3
For the Large BioMed matching tasks, the matcher is capable of aligning the small fragments within the given time frame of 6 hours. While ALOD2Vec Matcher performs slightly above the 2017 and 2018 F1 averages on the small FMA-NCI data set, it perfoms in the lower half for the remaining ones. 2.4
Although the matcher presented here is not capable of generating complex correspondences yet, it could produce results for the entity identi cation subtask for two data sets: On GeoLink, ALOD2Vec Matcher achieved the highest F1 score and recall of all matchers that participated; on Hydrograph, alignments for the English ontologies could be generated and scored within the median.
The matcher performs above the given baselines. However, the matches are still rather trivial and mostly share common tokens.
There are multiple reasons for the mediocre performance. First, the underlying data set is very noisy: It contains a lot of wrong information (e.g. sh skos:broader sher )7, subjective information (e.g. donald trump skos:broader lunatic)8, and is not strictly hierarchical (e.g. live skos:broader quality, and vice versa)9. In addition, the tail-entity problem is still not solved because very speci c entities are involved in very few hypernymy statements and their resulting vectors are likely not meaningful (e.g. complex congenital heart defect )10.
Besides the pitfalls of the data set, the matcher cannot handle homonyms, non-nouns, or non-English labels. 3.2
There are three ways in which the current research focusing on this approach can be improved in the future: Firstly, more propositionalization techniques for very large data sets could be explored. Secondly, the matcher itself can be enhanced to use more information available in ontologies such as their structure. And lastly, the data sets to be used can be improved. WebIsALOD is only one Web-scale RDF data set and still has some pitfalls such as the restriction to hypernymy relations and noise. More such data sets can be created and used in the future. 4
In this paper, we presented the ALOD2Vec Matcher, a matcher utilizing a Webcrawled knowledge data set by applying the RDF2Vec methodology to a hypernymy data set extracted from the Web. It could be shown that it is possible to use very large RDF graphs as external background knowledge and the RDF2Vec methodology for the task of ontology matching.
8 see http://webisa.webdatacommons.org/concept/donald_trump_ 9 see http://webisa.webdatacommons.org/concept/_quality_ 10 see http://webisa.webdatacommons.org/concept/complex+congenital+heart_ defect_