The wides pread of semantic web technologies such as RDF, SPARQL and OWL enables individuals to build their databases on the web, write vocabularies, and de ne rules to arrange and explain the relationships between data according to the Linked Data principles. As a consequence, a large amount of structured and interlinked data is being generated daily. A close examination of the quality of this data could be very critical, especially if important researches and professional decisions depend on it. Several linked data quality metrics have been proposed, and they cover numerous dimensions of linked data quality such as completeness, consistency, conciseness and interlinking. In this work, we are interested in linked data quality dimensions, especially the completeness and conciseness of linked datasets. A set of experiments were conducted on a real-world dataset (DBpedia) to evaluate our proposed approaches.
Because a large amount of information is being generated daily, and
information needs to be of high quality to be useful, the need for quality assessment of
this data on the internet is more urgent ever before. On the other hand, Linked
Open Data 1 (LOD) has appeared as a result of the development of semantic
web technologies, such as RDF, SPARQL and OWL. A research of the quality of
information has been successfully applied on the traditional information system,
with the rational databases having a positive impact on the used organizational
processes [
1 http://5stardata.info/en/
Completeness is a data quality measure that refers to the degree to which
all required information is present in a particular dataset [
Scientist Schema = fProperties on Scientist g [ fProperties on Persong [ fProperties on Agent g [ fProperties on Thing g
such that: Scientist v P erson v Agent v T hing
However, we can obtain the size of Scientist Schema, which is equal to 664 (ABox properties) in the case of DBpedia with a simple SPARQL query3. Thus, the completeness of the description of Albert Einstein could be calculated as follows:
Although, the property weapon is in Scientist Schema, but it is not relevant
for the Albert Einstein instance. As a conclusion, we can nally say that the
completeness as calculated here does not provide us with the relevant value regarding
the real representation of scientists in the DBpedia dataset. Hence, we need to
overcome this issue by inventing and exploring to get an idea about how they
are actually described and which properties are used. In [
Our research on conciseness dimension was inspired by the existing Synonym
Analysis for Predicate Expansion [
Range content ltering aims to represent a transaction as a distinct object with several predicates. For example, the object Lyon city is connected with several predicates such as (birthPlace, deathPlace and location). The authors suppose that synonym predicates have a similar sense. They also share a similar group of object values. For this reason, the proposed approach nds that the frequent sets pattern of predicates is dominated by object values.
Thus, it is not su cient to discover the predicates that are used synonymously depending on Range Content Filtering alone. For example, the predicates birthPlace and deathPlace share the signi cant co-occurrences with the same object values but they are de nitely used di erently. However, the authors have proposed another lter in order to overcome this problem and to nd the synonym predicates more correctly. They expect that the synonym predicates should not co-exist for the same instance. According to schema analysis, transactions of distinct subjects with several predicates are represented. By applying negative association rules, the synonym predicates appear in di erent transactions. For instance, the subject Michael Schumacher does not have two synonymously used predicates such as born and birthPlace in the same dataset.
Now, our objective is to discover synonym predicates by applying the proposed approach. We clarify its drawbacks through applying the next example (see Table 1), and we would like to apply the previous approaches on a sample of facts from DBpedia to discover the synonym predicates.
Based on range content ltering, all predicates will be gathered into groups
by each distinct object. Thus, results can be as illustrated in Table 2 in order to
retrieve the frequent candidates. As a result, we can see that the nationality and
sourceCountry predicates are already in the same transaction. By applying
FPgrowth algorithm [
SAAAWWWudddhhhbaaaiiijmmmttteeeectHHHRRRaaaiiivvvdddeeewwwrrriiinnn tbrssnPyitoiarrvapuetteteidrherociPMnceaCalatoleoicutuetynhtry LCCOOGMaboaanoklnnjtofeaeaascPSrddetiulaaJoapayewerrior LCOOGMaboanoklnjtofeaeascPSrdetiulaJoapayewerrior tbrsnPyitiarrvapetteteidheroiPMncaalatoleicutetyh, sourceCountry Table 1: Facts in SPO structure Table 2: Range Content Filterfrom DBpedia ing Subject Predicate Adam Hadwin type, birthPlace , nationality White River sourceCountry,riverMouth,state
Table 3: Schema analysis
Therefore, the algorithm proposed the nationality and sourceCountry as synonym predicate pairs, which is not correct because we cannot replace nationality predicate that is related to Person class as its Domain with sourceCountry predicate, which is related to Stream class as its Domain.
Nowadays we are witnessing an increase in data accessible on the internet.
There are large amounts of data being generated daily. It plays a crucial role
in companies, organization and individual decisions. This data, although rich in
content, is often incomplete, lacks metadata or even su ers from redundancy. As
our goal is to improve Linked Data quality, the problem is relevant for Linked
Data publishers, contributors and consumers. Users look forward to getting
information with a high quality which means that data is \ tness to use" [
Several metrics and tools have been proposed to assess Linked Data and
improve its quality [
Completeness is one of the essential dimensions of data quality, which refers to
the amount of the presented information. Pipino et al. [
On the other hand, Mendes et al. [
Our goal is to enhance the dimensions of linked data quality that do not have enough metrics (i.e. completeness and conciseness). We aim to propose new metrics from di erent perspectives such as inferring a reference schema from data source and using semantic analysis to understand the meaning of predicates.
Completeness calculation requires a reference schema to be compared with. The gold-standard or prede ned schema does not always represent a good reference. So, there is a need to explore instances to have a suitable reference schema (ontology). Also, dataset ontology contains semantic features which represent an explanation of the meaning of each predicate.
{ Completeness dimension: Is it possible to calculate completeness values using inferred schema from data source? How can we assess the completeness of Linked Data? { Conciseness dimension: Can we enhance the conciseness dimension of liked datasets by analyzing the semantics of predicates? 5
Our hypotheses are derived directly from the questions above:
H1 Exploring instances to get an idea about how they are actually described and which properties, besides considering the importance of each one, are used. This provides more suitable schema to use as a reference one in order to calculate completeness value of a dataset.
H2 A deep semantic analysis of data, beside to the statistical analysis, can enhance the conciseness of linked datasets by discovering repeated predicates. Where the semantic analysis will reduce the false positive results.
Completeness assessment
On the basis of our belief that a suitable schema (e.g. a set of properties) needs to be inferred from the data source, the experiments were performed on the well-known real-world datasets, DBpedia, publicly available on the Linked Open Data (LOD). DBpedia, is a large knowledge base composed of structured information extracted collaboratively from Wikipedia. It describes currently about 14 million things.
In [
The experiments revealed that datasets completeness could increase or decrease due to changes made to existing data or to the new added data. We also noticed that often this evolution does not bene t from the initial data cleaning as the set of properties continue evolving over time. Our approach could be helpful for data source providers to improve, or at least to keep a certain completeness of their datasets over di erent versions. It could be particularly useful for datasets constructed collaboratively by applying some rules for contributors when they update or add new resources. 6.2
The automatic generation of RDF knowledge bases might lead to several semantic and syntactic errors in addition to incomplete metadata.
Because publisher commonly do not respect the semantics including misuses
of ontological term and unde ned classes and properties [
We represent, in this section, our approach [
Our mining-based approach includes two steps:
1. Properties mining: Given a dataset D, we rst represent the properties,
used for the description of the D instances, as a transaction vector. We then
apply the well-known FP-growth algorithm [
De nition 1. ( Completeness) Let I0 a subset of instances, T the set of transactions constructed from I0, and MF P a set of maximal frequent pattern. The completeness of I0 corresponds to the completeness of its transaction vector T obtained by calculating the average of the completeness of T regarding each pattern in MF P. Therefore, we de ne the completeness CP of a subset of instance I0 as follows:
CP(I0) = 1 jT j jMFPj (E(tk); P^j )
X X jT j k=1 j=1
jMF Pj such that: P^j 2 MF P , and (E(tk); P^j ) = n 1 if P^j E(tk) 0 otherwise (1) 7.2
The objective of the semantic analysis is to nd the meaning of the predicate. Depending on systematic analysis alone is not su cient to discover the synonymously used predicates, also too many false positive results are represented, especially when we deal with a large dataset. As the previous example illustrated in Section 1.2, the predicates nationality and sourceCountry can have the same object like Canada. They also never appear or co-occur together for the same subject. Obviously, the nationality is a predicate of Person class and sourceCountry is a predicate of Stream class.
We add an important extension to the previous work by studying the meaning of each candidate. In addition, we study some conditions to examine them exploring their meanings so that we mathematically prove on a basis of Description Logic that a predicate cannot be a synonym of another predicate if they have disjoint domains or ranges. Through taking the same previous example of nationality and sourceCountry predicates, we will analyze the domain and range of each one of them. On one hand, the predicate nationality has a domain as Person class and a range as Country class, and on the other hand, the predicate sourceCountry has a domain as Stream class and a range as Country class. According to DBpedia ontology Stream class is a subclass of Place class, as well as Place and Person classes are completely disjointed. Consequently, we cannot consider nationality and sourceCountry as synonym predicates.
To promote our arguments, we will prove that a predicate cannot be a synonym of other predicate in some cases according to the semantic features of each one, such as: Disjoint properties based on their domains and ranges, Symmetric/Asymmetric Property, Inverse Functional property, Functional property and max cardinality. We illustrate these arguments using Description Logic formalization.
Our goal is to compare the completeness and conciseness of dataset with our
presented approach to the state of the art such as Sieve [
Since we believe that poor-quality data a ects negatively on the decision that can lead to catastrophic consequences, improving the quality of linked data is our research main aim because Web of Data is worthlessness without good quality. We are concerned to concentrate on two dimensions, which do not have a lot of metrics or indications to be evaluated, according to what Zaveri suggested. We believe that extracting a reference schema from data source is more suitable to calculate the completeness of dataset, beside, we prove the importance of semantic part in addition to the statistical one to enhance the conciseness of dataset. Our proposed approach takes into account only properties disjoint and functional proprieties because semantic features are not su ciently present as explained in Section 6.2. Therefore, our plan is to infer all possible semantic features of LOD datasets too.
Acknowledgements I would like to thank my supervisors Prof. Dr. Samira Si-said Cher and Dr. Faycal Hamdi for their support and the opportunity for the realization of this work.