When ontologies cover overlapping topics, the overlap can be represented using ontology alignments. These alignments need to be continuously adapted to changing ontologies. Especially for large ontologies this is a costly task often consisting of manual work. Finding changes that do not lead to an adaption of the alignment can potentially make this process signi cantly easier. This work presents an approach to nding these changes based on RDF embeddings and common classi cation techniques. To examine the feasibility of this approach, an evaluation on a real-world dataset is presented. In this evaluation, the best classi ers reached a precision of 0.8.
Finding alignments between ontologies, also known as ontology matching, is a
non-trivial task and has been an active area of research over the last ten years.
Several approaches in this area are based on the structure of the ontologies,
logical axioms or lexical similarity [
In this paper, we propose an approach to this problem based on RDF embeddings and well-known classi cation techniques. The central aspect of this approach is to represent changed concepts by their RDF embedding and classify whether an alignment statement nearby should be changed. To gain evidence if this approach works, we evaluate it using a dataset from the area of biomedical ontologies. On this dataset, our approach is able to identify changes a ecting alignment statements with a precision of 0.8.
The remainder of this work is structured as follows: Section 2 discusses foundations of our work and related approaches. The general approach is presented in Section 3. Evaluation methodology, the dataset and results are shown in Section 4. Section 5 discusses the results of our evaluation, and advantages and disadvantages of our approach. A conclusion is given in Section 6. 2
An ontology alignment (sometimes also called ontology mapping) is a set of
correspondences between entities in di erent ontologies [
AO1;O2 = f(c1; c2; semT ype)jc1 2 O1; c2 2 O2; semT ype 2 f ; ; gg AO1;O2 is the set of all alignment statements. To denote a change of an ontology over time, we use the prime symbol (e.g., a changed version of O is denoted as O0). The alignment adaption problem for two ontologies O1 and O2 connected by AO1;O2 can then be stated as nding a new alignment A0O10;O20 , when O1 and O2 evolve to O10 and O20.
In the area of ontology alignment adaption, several approaches are based on
rules or rule-based dependency analysis. [
The task of knowledge base completion shares some properties with the
problem we address in this work. In that area, classi ers are given a subject and a
predicate and try to predict an object [
To our knowledge, no approach exists that predicts whether a given change has an impact on the alignment without using a detailed set of rules. This issue is at the core of our research.
Our general approach is based on the representation of changed resources using
RDF embeddings, a represenation of RDF nodes as vectors in a high-dimensional,
dense vector space. RDF embeddings are generated using RDF2Vec [
Each changed resource is represented by the corresponding RDF2Vec vector. Hence, the input to the training of the classi er is a pair (v(c); k) consisting of a vector v(c) and a class k. k determines whether c caused a change in its direct neighbourhood. The task at hand is to correctly classify new vectors. To solve this problem, we use several common classi cation techniques: Regression, Naive Bayes, Tree-Based Algorithms as well as Support Vector Machines and Multilayer Perceptrons. Each algorithm is trained on one set of changes and evaluated on a di erent set. 4
The research questions behind our evaluation are the following: 1. Can RDF embeddings be used for change classi cation with an acceptable performance? This question tries to clarify, whether our approach is in general applicable to the problem at hand. 2. Which classi ers can be used for this problem? This question is used to identify the best classi ers for our problem. 4.1
The dataset used to answer these research questions in our experiments is a real-word dataset from the domain of biomedical ontologies. It has been used
in several works that deal with alignment adaption, e.g., [
For simplicity of our presentation, we will only present the alignment between the ontologies NCI and FMA in the version change from 2009 to 2010. In the formal notation introduced in Section 2, O1 and O2 refer to the ontologies NCI and FMA as of 2009 and O10 and O20 as of 2010, respectively. The alignment from 2009 is denoted by AO1;O2 and the version from 2010 by A0O10;O20 .
From 2009 to 2010, 924 changes are near alignment statements of which 47% require an adaption. These changes are used as a training set. The test set consists of the changes from 2010 to 2011. This set contains 785 changes near alignment statements, of which 36% lead to an alignment adaption. 4.2
To generate RDF embeddings, the code from RDF2Vec [
ontologies_and_mappings/ontology_mapping_adaption 3 http://scikit-learn.org/stable/
The results of the described process are displayed in Table 2. Only changes close to the alignment were included in this evaluation, since it would otherwise be very easy to achieve accuracy values above 95%. Results for MLP did not vary based on the structure of the hidden layers, so one row represents all MLP results. All algorithms show a very similar performance regarding the evaluated metrics. The highest achieved precision is 0.81, which can be reached using MLP and linear SVM classi cation. These methods also reach the highest f1-measures of 0.75. Accuracy of all classi ers is only marginally higher than what can be achieved using random guessing, given the distribution of classes in the test set. The results presented in section 4.3 give us some evidence on our rst research question: Using RDF embeddings to represent changes seems to be a promising approach to the mapping adaption problem, as we can see a precision around 0.8. In general, several classi cation approaches show a similar performance. This precision can be achieved, although the approach uses no information regarding the nature of changes, e.g., the algorithm can not distinguish the correction of typos from major, structural changes.
An important advantage of this approach is that no sophisticated change
model that is adapted to the domain is required. Approaches like [
In this work, we presented an approach to ontology alignment adaption based on RDF embeddings and common classi cation techniques. An evaluation on a dataset from the biomedical domain provided some evidence, that the approach is feasible. On the dataset, best-performing classi ers had a precision of 0.8.
As future work, several extensions are possible: Further evaluations could be performed on di erent datasets. Also, a combination of this approach with existing mapping adaption approaches could be examined. Change types could be used as another input to the classi cation process to improve classi cation accuracy. Another aspect for future work is to determine, when embeddings need to be updated, since embeddingis will become outdated when ontologies change.