Keywords

1 Introduction

Linked Open DataFootnote 1 (LOD) is an initiative proposed by Tim Berners-Lee to publish data on the Web. Such datasets use standards from the Semantic Web (SW) like the Resource Description FrameworkFootnote 2 (RDF) to represent data. RDF is an oriented and labeled multi-graph model. LOD datasets are graphs and thus can be linked together through the addition of statements having one node in one dataset and another node in a different dataset. This is the linking part of Linked Open Data.

There are more and more LOD datasets published and linked together. Linking datasets is a powerful way to enable retrieving knowledge from other datasets. This enables the ability of crawl various datasets to discover new facts. Thanks to ontologies using OWL [16], it is even possible to use reasoners based on, e.g., Description Logics (DL), to infer new things from data.

There are many ways to link two datasets. Many properties can be used to do so, since virtually any object propertyFootnote 3 can be used. One of the most employed is the famous owl:sameAs property from the OWL ontology. To create a link with owl:sameAs between two instances is stating that these two instances are the same. It is an identity link. For example, let us say that we have two instances ds1:Paris and ds2:CityOfParis of Paris, the French capital. Each one of these instances is in one dataset (ds1 and ds2 here). It is possible to link ds1 and ds2 by simply adding this fact: owl:sameAs(ds1:Paris, ds2:CityOfParis).

Because owl:sameAs has a very strict semantics, creating such links among instances has consequences. owl:sameAs semantics is based upon Leibniz’s principle of indiscernibility. Therefore if two things are identical they must share the same values for the same properties. If in ds1 the population of Paris is not filled out but is given in ds2, then the population can be used in ds1. More formally, the indiscernibility of identicals is: \(owl{:}sameAs(a, b)\rightarrow (p(a,o) \rightarrow p(b,o))\). This is how one can discover new knowledge through the use of owl:sameAs links, by either retrieving or inferring new data. An erroneous link can lead to inaccurate inferred data and consequently it is very important to have high quality owl:sameAs links in order to maintain the overall quality of the datasets. As stated in Ding et al. [6] and Halpin et al. [12], there are many misuses of owl:sameAs in current LOD, which undermines the quality of the data it provides.

2 State of the Art

There are many facets to the interlinking problem. We want to underline four of them in this section.

2.1 Instance Matching

The term instance matching refers to the problem of finding equivalent resources. As stated in Hogan et al. [15], the search for identity links among instances of the LOD has several names in literature such as data linking [10], data reconciliation [25], record linkage [21], duplicate identification [7], object consolidation [14], instance matching [3], link discovery, or Co-reference resolution. Those approaches are historically the first ones to emerge. The goal is to produce links between a source dataset and a target dataset. For each potential link a similarity score is produced and if the score is above some threshold, the link is validated. Ferraram et al. [8] published a complete survey and more recently Achichi et al. [1] and Nentwig et al. [20] propose complementary surveys.

2.2 Knowledge Enrichment

Because several approaches used to match instances benefit from semantics features like inverse functional properties or cardinality restrictions, knowledge enrichment approaches are tangential to our domain of concern. For example, if one can find that a property is inverse functional, then any subjects related to the same object by this property have to be the same. More formally: \(InverseFunctional(P)\wedge P(a,c)\wedge P(b,c)\rightarrow owl{:}sameAs(a, b)\). Thus, the addition of new knowledge to the TBoxFootnote 4 of an ontology might, hypothetically, lead to finding novel owl:sameAs links. Völker and Niepert [27] propose the induction of a schema (i.e. an ontology) for a given KB by using statistics. Association rules are mined and then translated into OWL2 axioms, but some interesting features cannot be mined, e.g. inverse properties, cardinal restrictions or property disjointness. Töpper et al. [26] also use statistical methods but from the Inductive Logical Programming (ILP) field, where only property domain, property range and class disjointness are computed. With AMIE [9], the authors use an assumption called the partial completeness assumption (PCL) and calculate scored rules under this assumption. PCL assumes that if a subject-predicate pair is present in the KB, then all possible objects for that pair are in the dataset. This method does not take into account any existing ontological knowledge nor use reasoning capabilities of Description Logics (DL). The approach of d’Amato et al. [4] is able to use ontological knowledge to produce rules.

2.3 Identity Crisis

Since several works, like Halpin et al. [12] or Ding et al. [6], raised the misuse of identity links, proposals have been made to circumvent this problem. Halpin et al. [12] propose to use weaker versions of owl:sameAs (e.g. from SKOS vocabulary) but with a loss of inference capabilities. There are also proposals for new properties to represent identity relations by McCusker and McGuinness [19] and Halpin et al. [12], where owl:sameAs is a sub property of a more specific and more relaxed list of (semi-)identity properties. [12] by using a new ontology where owl:sameAs is a sub property of more specific and more relaxed link of (semi-)identity links. In this hierarchy each property has a combination of reflexivity, symmetry and/or transitivity. McCusker and McGuinness [19] propose a local and domain-specific approach where one can specialize owl:sameAs property in a particular domain (e.g. using a biomedidentity:sameAsBioSource property in the biology domain) but link usability thus becomes only local. At the intersection of link invalidation and contextual identity links, De Melo [5] propose to use a property to assert proven identity links on the basis that owl:sameAs might contain erroneous links. That is if an owl:sameAs link between a and b is proven to be right then one can have a lvont:strictlySameAs link between a and b. In Halpin et al. [13], authors propose to manage the context of identity links through the addition of a formal context. Several ideas are proposed, but none of them has been widely adopted. Beek et al. [2] propose to manage the context as a set of properties. Therefore two things are equal if they share the same property-value pairs where properties are defined by the context. Idrissou et al. [17] extends the proposition of Beek et al. [2] by adding operators other than the intersection between the sets of properties representing different contexts. Raad et al. [24], also based on Beek et al. [2] work, propose an algorithm to compute those contexts, i.e. contexts based on sets of properties where identity holds.

2.4 Identity Link Assessment

Identity link assessment approaches consist of checking if a link is true or false. It does not create any link but does evaluate existing ones. Guéret et al. [11] propose to use classical network measures to assess existing links. De Melo [5] propose an approach using the unique name assumption within datasets (i.e. an instance has one name in a dataset) to spot sets of instances linked by owl:sameAs where at least one link is presumed to be wrong. Next, a linear programming algorithm is used to compute the wrong link. In Papaleo et al. [22], the authors propose a logical approach to detect such wrong statements. The algorithm tries to detect logical conflicts by using semantics features like functional properties in small sub-graph containing the two involved instances to assess. Thus, this approach strongly relies on semantics. Paulheim [23] propose to use data mining methods. First, links are represented in an embedded space. Second, an outlier detection algorithm is used to detect links that may be erroneous.

3 Problem Statement and Contributions

As we have seen in Sect. 1, a wrong owl:sameAs link can lead to inferring wrong data and therefore reduce the overall quality of datasets. As stated by Halpin et al. [12] there is an identity crisis in the sense that the owl:sameAs strict semantics is not always respected when used. To state that two things are the same is a very strong statement in LOD domain, with huge consequences. It is common to find instances that are nearly the same but not quite. For example, the city of Paris is geographically the same as the administrative department of Paris (it is an administrative subdivision). But in a legal context they are two completely different things. Another example is a glass of water belonging to a set of glasses. All glasses look the same and in most contexts they certainly are interchangeable but two glasses of this set are two physically different objects. What is identical or not is a philosophical question.

So there is confusion regarding how the owl:sameAs property has been defined and how it has been used. owl:sameAs has been created to be used only when things are really the same, in all contexts. But in the wild it is used in a more relaxed context, leading users to infer inaccurate facts.

There is a need to know how incorrect owl:sameAs links decrease the overall quality, and how frequent such incorrect links are. Also it is important to investigate the use of semantic features among LOD datasets. Because owl:sameAs links computation also relies on semantics features (e.g. functional properties, maximum cardinality, etc.).

Our goal is to identify quality defects related to data interlinking, then to assess them and finally, propose a way to correct them. Some questions that this work proposes to address include: How to evaluate a owl:sameAs link? What is the impact of erroneous links of this type on datasets? How to measure this impact and how to reduce the negative effect on dataset quality?

4 Research Methodology and Approach

For now, the approach we consider is the following, depending on preliminary results:

  1. 1.

    The first step consists in identifying quality defects due to data interlinking. As pointed out by several authors, owl:sameAs links are far from sure things. They can be incorrect and therefore produce incoherence in datasets. Thus, we propose to investigate several domains and try to find the nature of the underlying quality defects such as incompleteness (missing links) or incoherence. We will have to define efficient algorithms for data exploration and defect detection.

    But before before diving deeper into quality defects, we plan to reproduce independently some results from previous work (like in Halpin et al. [12]) and to gather useful and precise statistics about owl:sameAs links and semantics in general. By semantics we mean OWL properties or classes, because owl:sameAs links rely a lot on them. Moreover we need to establish how much this identity crisis is a problem.

  2. 2.

    For each type of defect, we need to associate a set of assessment methods and algorithms. The proposal of Papaleo et al. [22] relies mainly on semantic features like functional or inverse functional properties. Problems arise when semantic features are not sufficiently present in the vocabulary or the data. For example, according to the last version of DBpedia (October-2016), the ontology contains only 30 functional properties (1%) and zero inverse functional property (out of more than 2860 properties). 1.6M out of 4.6M instances (34%) use at least one of the functional properties. Conversely, other approaches rely on statistical, graph or mining techniques. A hybrid approach using semantics, statistics on data and vocabulary, network or graphs measures (etc.) could be used successfully.

  3. 3.

    The next logical step is to be able to correct detected quality defects. Such defects decrease data quality and hence, finding how to correct them can improve data quality. There are, at this time, not so many propositions to deal with existing defects, where the balance tends to lean towards the creation of links in the literature.

  4. 4.

    We want to implement the proposed solutions through a Web application and/or a Web service to support Data Interlinking evaluation and improvement. This is an important point because if the tooling is not good enough, we cannot expect people to adopt our work.

  5. 5.

    Finally, with the high volume of data composing the LOD cloudFootnote 5, it is not only a matter of Open Data but also a Big Data issue. Hence, the developed solutions should scale well.

5 Preliminary or Intermediate Results

Until now, we have focused primarily on related work. Nevertheless, we have written two articles to study some ideas related to our problem.

One of the interesting aspects of linking datasets is that one can complete data in a dataset thanks to another one. Therefore, studying completeness might be interesting because to understand the impact of interlinking between two datasets one has to know more about completeness of datasets. Once two instances are connected thanks to an owl:sameAs link, one can retrieve properties from the first instance towards the second one (and conversely). As a consequence, completeness might be subject to variation and to be able to assess this variation before and after interlinking is important information. We want to be able to check how the owl:sameAs property, used to connect a local dataset to external datasets, can help improve the completeness of this given local dataset. Furthermore, even if it might not be obvious at the first look, completeness is also a very important part to interlink two datasets, i.e. to produce links. In fact, the more complete the datasets, the easier it is to find proof that two instances are the same. For example, with only the last name we can not tell if two persons are the same or not. But in addition with their first names and birth dates, we have more hints to decide whether they are the same or not. Hence, a method for assessing completeness can be very interesting given the two previous points. We published a study [18] to assess the completeness evolution of DBpedia. In this work we proposed to assess completeness by using a mining-based approach. Using a given set of instances (\(\mathcal I\)) of the same class, the first step consists in computing a set of properties that are frequently found among instances of \(\mathcal I\). For each instance, the corresponding transaction is all its properties. Here we use frequent-itemset algorithm to do so. Once those set of properties (or transactions) found, we can compute the completeness \(\mathcal {CP}\) of the set of instances \(\mathcal I\):

$$\begin{aligned} \mathcal {CP}(\mathcal I)=\frac{1}{|\mathcal T|}\sum _{k=1}^{|\mathcal T|} \sum _{j=1}^{|\mathcal {MFP}|} \frac{\delta (\mathcal {P}(t_{k}),\hat{P}_j)}{|\mathcal {MFP}|} \end{aligned}$$
(1)

such that \(\mathcal T\) is the set of transactions associated with \(\mathcal I\) (see second column of Table 2). \(\mathcal {MFP}\) is the set of maximal frequent patterns computed in the first step, \(\mathcal {P}(t_k)\) is the power set of the transaction \(t_k\) and \(\delta \) is equal to one if the frequent pattern \(\hat{P}_j\) is in \(\mathcal {P}(t_k)\), zero otherwise.

Table 1. A sample of DBpedia triples
Full size table
Table 2. Transactions corresponding to triples from Table 1
Full size table

Example 1

The completeness of the subset of instances in Table 1 regarding their transactions in Table 2 and \(\mathcal {MFP}=\) {{director,musicComposer}, {director, editing}}, would be:

$$\begin{aligned} \mathcal {CP}(\mathcal I')=(2*(1/2)+(2/2))/3=0.67 \end{aligned}$$

This value corresponds to the completeness average value for the whole dataset regarding the inferred patterns in \(\mathcal {MFP}\).

We apply this approach on several versions of DBpedia, thus allowing us to study evolution when new data is added or changed. This permits us to have an efficient way to characterize an important aspect of data quality, i.e. completeness.

We are also investigating the way owl:sameAs links are created. We are currently writing a proposition that takes into account the structure of data (both TBox and ABoxFootnote 6) in a dataset. We believe that not all evidence (in one direction or another) have the same strength. To compare two instances from two distinct datasets we want to collect evidences. Evidence is either a proof of identity or a proof of difference between those two instances. If the two instances share a property and an object, it is therefore a proof they may be the same. Rather, if they share a property but the objects are different, this evidence will tend to prove that the instances are different. The strength of this proof depends on the weight of the property among instances of the same class, and it also depends on the discriminating power of the object(s). The higher the weight and discriminating power are, the stronger the evidence. This approach can help us determine whether an existing relationship has a justification.

6 Evaluation Plan

The rudimentary version of our evaluation plan is as follows:

  1. 1.

    In order to validate our work, we will first select several interlinked datasets from the LOD. Datasets will have to cover different domains, with different sizes and vocabularies to ensure a large number of situations to be tested. For the moment, DBpedia and Wikidata are good candidates to be used as a playground since both datasets are cross-domain, well known and well interconnected. Datasets in a specific domain should also be part of our assessment. Therefore, we also plan to use life sciences datasets, as there has been an explosion in the number of interlinked datasets in this area. Datasets from other domains can be used later if necessary.

  2. 2.

    For each dataset, we will have to evaluate several dimensions of data quality (still to be chosen).

  3. 3.

    The third step is to apply our approach. We will assess the interlinked datasets and modify them based on the experience gained in previous steps (e.g., to help determine which links can be safely deleted, which are useful in one context but not in another, etc.). This step will depend in large part on our future findings, so it may be subject to change.

  4. 4.

    Finally, we will again compute the same data quality dimensions. At this stage, we will have assessed several dimensions of data quality before and after the application of our approach, and so we can compare the situation before and after, hoping that there will be an improvement.

7 Conclusions

Being able to identify inaccurate links is a first step towards improving the overall quality of the data, since the data producer could be notified at the time of publication. Moreover, since the correction of links is not sufficiently addressed in the literature, if we achieve this objective, improving existing data sets would be a significant breakthrough in the identity crisis of the LOD.