Keywords

1 Introduction

Since the introduction of the Linked Open Data (LOD) cloud, the general purpose KGs like DBpedia, YAGO, Wikidata have been the focal point of research in the field of data mining and information retrieval. Hence, the correctness and completeness of such KGs is of great importance. However, many studies show that information in these KGs often can be noisy, incorrect and incomplete [3, 6, 7, 9]. One way to account for the incompleteness of information in a KG is to harness the complementary information from different KGs.

Nevertheless, the different KGs are following different knowledge organization approaches [2, 4, 8] and use different underlying ontologies to represent knowledge, where explicit alignments amongst the different ontologies are not always available [5]. Therefore, a direct comparison of the KGs in the content level is a challenging task. For example, in Wikidata, the property wdt:P31 (instance of)Footnote 1 defines what we know as rdf:type. However, based on our observations, wdt:P31 follows different semantics and it differs in its use when compared to rdf:type in DBpedia. Thus, by relying only on wdt:P31 it is not possible to have a direct content-based comparison of the classes of the two KGs.

In this paper, we propose a light-weight isomorphism-based schema matching approach to harmonize two KGs having different underlying schema structure. For this study, we have used the two most popular KGs: DBpedia (English language) and Wikidata. The main aim of this work is to infer type subsumption relations in Wikidata by leveraging the existing equivalence relations between Wikidata and DBpedia. To this purpose we establish conditional subsumption relations between Wikidata properties and rdf:type .

2 Type Subsumption

Problem Description - We consider two RDFSFootnote 2 KGs, a source \(K_{S}\) and a target \(K_{T}\), consisting of set of triples \(K \subseteq E \times R \times (E \cup L)\), where E is a set of resources referred to as entities, L a set of literals, and R a set of relations. \(\{C_{S_{i}}\}\) and \(\{C_{T_{j}}\}\) is the set of classes in the source and target KG respectively. We assume that the classes and the entities of \(K_{S}\) and \(K_{T}\) are aligned i.e. \(K_{S}\) stores the statement \({<}C_{S_{n}}, \texttt {owl:equivalentClass}, C_{T_{m}}{>}\) and \({<}e_{S}, \texttt {owl:sameAs}, e_{T}{>}\).

In this work, we aim for a conditional subsumption relation alignment, as the schemas used for KGs vary heavily. Thus only equivalence alignments that have merely similar semantics or subsume one another are not enough to map the relations. Following the relation subsumption definition in [5] the goal is:

Goal. For two KGs, a source \(K_S \subseteq E_S \times R_S \times (E_S \cup L_S)\) and a target \(K_T \subseteq E_T \times R_T \times (E_T \cup L_T)\), and a relation rdf:type \(\in \) \(R_S\), find relations \(r_T \in R_T\) s.t. \(r_T \subseteq \) rdf:type. The equivalence relation between \(r_{S}\) and \(r_{T}\), can also be expressed as a two-way subsumption relation: \(r_{S} \equiv r_{T}\), iff \(r_{S} \subseteq r_{T}\) and \(r_{T} \subseteq r_{S}\).

Methodology - The aforementioned goal is achieved by exploiting the equivalence relations of classes and instances between the two KGs. The method is described with the help of the illustration in Fig. 1.

  1. Step 1:

    For each class \(C_{S_{i}}\) in DBpedia, we determine the entities \(e_S\) of the class via rdf:type relation. Formally: \(\forall C_{S_{i}} \in K_S: {<}e_S, \texttt {rdf:type}, C_{S_{i}}{>}\)

  2. Step 2:

    From the entities \(e_S\), find those with owl:sameAs link(s) to corresponding \(e_T\) entities in Wikidata. Formally: \(\forall e_S \in C_{S_{i}}, \exists e_T \in K_T : {<}e_S, \texttt {owl:sameAs}, e_T{>}\)

  3. Step 3:

    Determine the class \(C_{T_{j}}\) in Wikidata equivalent to DBpedia class, \(C_{S_{i}}\) via the owl:equivalentClass relation. Formally: \(\forall C_{S_{i}} \,{\in }\, K_S, \exists C_{T_{j}} \,{\in }\, K_T: {<}C_{S_{i}}, \texttt {owl:equivalentClass}, C_{T_{j}}{>}5\)

  4. Step 4:

    For each entity \(e_{T_{j}}\), check if there is any relation (or relations) \(r_{T_{j}}\), which connects to \(C_{T_{j}}\). Formally: \(\forall e_T, \exists r_{T_{j}} \in K_T: {<}e_T, r_{T_{j}}, C_{T_{j}}{>}\)

Fig. 1.
figure 1

Isomorphic approach to infer type subsumption relations in Wikidata with the help of DBpedia

Full size image
Fig. 2.
figure 2

Comparison of the KGs (best viewed with color print) [1]

Full size image

3 Experimental Evaluation

This section discusses the results of the approach of inferring type subsumption relations in Wikidata leveraging existing mappings to DBpedia. Due to lack of space the full set of results can be found here [1].

For this work, all the experiments were carried out on DBpedia 2016-10 version and Wikidata as of January 11, 2018. Out of the 524 interlinked classes between DBpedia and Wikidata, we conducted experiments on 327 classes, the instances of which are linked via owl:sameAs.

Results - The experiments establish the fact that the type information in Wikidata is often implicitly defined and 41 properties, including wdt:P31 (instance of), hold a subsumption relation with rdf:type in DBpedia. Interestingly only the members of about 38% of these Wikidata classes can be accessed via wdt:P31. Furthermore, only 58% of the aforementioned 38% of Wikidata classes are using the property wdt:P31 exclusively to denote the membership in a class. Table 1 shows some Wikidata classes and the properties serving as rdf:type ordered by the percentage of the class members which were retrieved via them.

Additionally, it is also interesting to notice that similar classes have similar type subsumption relations. For instance, for the classes in Wikidata denoting different kinds of professions such as, Artist, Scientist the property occupation (wdt:P106) defines the members of the class.

Figure 2 illustrates a comparison between DBpedia and Wikidata for 5 classes. It is interesting to notice that the number of instances retrieved from Wikidata via the new type subsumption relations (red bar) is much higher than via only wdt:P31 (blue bar). Hence, more members of the classes can be retrieved using the subsumption relations leading to a strong foundation for the content level comparison of the KGs.

Furthermore, the green bar in the Fig. 2 represents the number of instances of the corresponding Wikidata classes using the type subsumption relations of Table 1, which also have owl:sameAs links to DBpedia. For all these classes, it has been observed that the height of the red bar (count of instances with new type subsumption relations) is higher than the green bar (count of instances with new type subsumption relations and owl:sameAs links to DBpedia), which reflects that Wikidata potentially contains more information than DBpedia for these classes. Also, it can be inferred that some of these entities in Wikidata are also present in DBpedia but are assigned to some other classes in DBpedia. This however can lead to further research on the correctness of the KG content.

Last, for the classes dbo:Animal and dbo:Plant, the number of instances in DBpedia (yellow) is higher than the number of instances that possess owl:sameAs links (green). Thus, some of the instances of these two classes in DBpedia are not instances of the corresponding owl:equivalentClass in Wikidata.

Table 1. Type subsumption properties for Wikidata classes [1]
Full size table

4 Conclusion and Future Work

This paper presented an isomorphic approach to infer type subsumption relations in Wikidata with the help of DBpedia. This approach can be extended to any two arbitrary KGs sharing equivalent classes and some equivalent instances. The results obtained in this study can be used as a starting point of further research on discovering potential errors or violations in the content of KGs. Next, we will explore the implicit type information stored in these KGs and contribute towards their completeness by predicting the type information using structural embeddings.