=Paper=
{{Paper
|id=Vol-2181/paper-05
|storemode=property
|title=Linked Data Quality
|pdfUrl=https://ceur-ws.org/Vol-2181/paper-05.pdf
|volume=Vol-2181
|authors=Subhi Issa
|dblpUrl=https://dblp.org/rec/conf/semweb/Issa18
}}
==Linked Data Quality==
https://ceur-ws.org/Vol-2181/paper-05.pdf
Linked Data Quality
Subhi Issa
Conservatoire National des Arts et Métiers, CEDRIC,
subhi.issa@cnam.fr
Abstract. 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 define rules to arrange and explain the re-
lationships 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 com-
pleteness, 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.
Keywords: LOD, Linked Data Quality, Completeness, Conciseness
1 Problem Statement
Because a large amount of information is being generated daily, and informa-
tion 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 [3]. This raises the question of the applicability of this approach in the
context of web of data. Zaveri et al. [17] surveyed 18 different linked data quality
dimensions that can be applied to assess the quality of Linked Data. The goal
of this work is to propose approaches that focus on figuring whether this infor-
mation completely represents the real world and that is logically consistent in
itself. Our objective is not measuring an absolute completeness and conciseness
but rather measuring their aspects.
1.1 Completeness
Completeness is a data quality measure that refers to the degree to which
all required information is present in a particular dataset [17]. We illustrate in
1
http://5stardata.info/en/
this section the main idea behind our approach through an example that shows
the issues and the difficulties encountered in the calculation of a dataset com-
pleteness. Let us consider the set of scientists described in the well-known open
linked dataset, DBpedia. We would like to calculate the completeness of a scien-
tist description (e.g. Albert Einstein), which will be the proportion of properties
used in the description of this scientist to the total number of properties in
Scientist Schema. In DBpedia, the Scientist 2 class has a list of 4 properties
(e.g. doctoralAdvisor ), but these properties are not the only ones used in the
description of a scientist (e.g. the birthdate property is not present in this list).
Indeed, the Scientist class has a super class called Person. So, the description
of a scientist may also take into account properties of the Scientist class and all
its ancestors.
Scientist Schema = {Properties on Scientist} ∪
{Properties on Person} ∪ {Properties on Agent} ∪
{Properties on Thing}
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 (A-
Box 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:
|Properties on Albert Einstein|
Comp(Albert Einstein) =
|Scientist Schema|
21
= = 3, 16%
664
Although, the property weapon is in Scientist Schema, but it is not relevant
for the Albert Einstein instance. As a conclusion, we can finally say that the com-
pleteness 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 [2], the authors propose
a new approach to compute the completeness of instances based on similar ones
in Wikidata. For each instance, they find the most frequent properties among
instances that have the same type, and find the percentage of missed proper-
ties to calculate the completeness. This approach sometimes does not work well
when the instance has several values of the property instance of as a class such
as Writer and Player.
1.2 Conciseness
Conciseness is one aspect of linked data quality dimension, which basically
aims to avoid repetition through elements having the same meaning with differ-
2
http://mappings.dbpedia.org/server/ontology/classes/
3
Performed on: http://dbpedia.org/sparql
ent identifiers or names. The eliminating of the synonymously used predicates
aims to optimize the dataset to speed up processing.
Our research on conciseness dimension was inspired by the existing Synonym
Analysis for Predicate Expansion [1]. However, Abedjan et Naumann proposed a
data-driven synonym discovery algorithm for a predicate expansion by applying
both schema analysis and range content filtering.
Range content filtering 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 finds that the
frequent sets pattern of predicates is dominated by object values.
Thus, it is not sufficient to discover the predicates that are used synonymously
depending on Range Content Filtering alone. For example, the predicates birth-
Place and deathPlace share the significant co-occurrences with the same object
values but they are definitely used differently. However, the authors have pro-
posed another filter in order to overcome this problem and to find 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 different 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 pro-
posed 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 filtering, 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 FP-
growth algorithm [8], frequent itemsets have been mined, thus nationality and
sourceCountry are the consequences. The next step is applying schema analysis
as a subject in a context and we will get the following transactions (see Table
3). We can notice that by applying negative association rules, there is no co-
occurrence between sourceCountry and nationality predicates.
Subject Predicate Object Object Predicate
Adam Hadwin type Golf P layer
Golf P layer type
Adam Hadwin birthP lace M ooseJ aw
M ooseJ aw birthP lace Subject Predicate
Adam Hadwin nationality Canada
Canada nationality, sourceCountry Adam Hadwin type, birthP lace , nationality
W hite River sourceCountry Canada
LakeS uperior riverM outh W hite River sourceCountry,riverM outh,state
W hite River riverM outh LakeS uperior
W hite River state Ontario Ontario state Table 3: Schema analysis
Table 1: Facts in SPO structure Table 2: Range Content Filter-
from DBpedia ing
Therefore, the algorithm proposed the nationality and sourceCountry as syn-
onym predicate pairs, which is not correct because we cannot replace nationality
predicate that is related to Person class as its Domain with sourceCountry pred-
icate, which is related to Stream class as its Domain.
2 Relevancy
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 suffers 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 in-
formation with a high quality which means that data is “fitness to use” [11].
3 Related work
Several metrics and tools have been proposed to assess Linked Data and
improve its quality [4,13,12]. Unfortunately, there were obstacles related to the
absence of a clear definition of the word “Quality” since it has different meaning
from a domain to another. However, data quality is commonly conceived to
suite our use so that it has several aspects or dimensions, such as accuracy,
completeness and interlinking. In 2016, Zaveri et al. [17] identified a set of 18
different data quality dimensions, each dimension has at least one indicator or
a metric to assess the given dimension. Some of the proposed approaches deal
with one dimension [7,15] or several dimensions [13,12].
Completeness is one of the essential dimensions of data quality, which refers to
the amount of the presented information. Pipino et al. [14] divided completeness
into: Schema completeness that is the degree to which classes and properties are
not missing in a schema, property completeness which is the extent of the missing
property values of a specific kind of property, and population completeness that
refers to the ratio of objects represented to real-world objects. Since several works
provide metrics for the three completeness classifications [13,5], their defined
metrics evaluate the completeness by comparing it with a predefined schema
that could not provide an accurate value of dataset completeness.
On the other hand, Mendes et al. [13] categorized conciseness dimension into
intensional and extensional conciseness. The first type, which is the intensional
conciseness, measures a number of unique dataset elements to the total number
of schema elements, thus this measurement is represented on the schema level.
In a similar manner but on the instance level, extensional conciseness measures
the number of unique objects to the real number of objects in the dataset. In
the similar sense but with another naming “uniqueness”, Füber et Hepp [5] de-
fined the elements of representation like classes, properties and objects. Their
definition suggested uniqueness of breadth at the schema level and uniqueness of
depth at the instance level. In [1], the authors proposed an algorithm to discover
the synonym predicates for query expansions. They depended on mining simi-
lar predicates according to their subjects and objects. However, their approach
works well when dealing with a dataset that has a limited number of instances.
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 different perspectives such as inferring a reference schema from data
source and using semantic analysis to understand the meaning of predicates.
4 Research questions
Completeness calculation requires a reference schema to be compared with.
The gold-standard or predefined schema does not always represent a good refer-
ence. 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 Hypotheses
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.
6 Preliminary results
6.1 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 infor-
mation extracted collaboratively from Wikipedia. It describes currently about
14 million things.
In [10], for evaluating the completeness of different versions of DBpedia,
we chose three relatively distant versions. The first one (v3.6) was generated
in March/April 2013, the second one (v2015-04) in February/March 2015 and
the third one (v2016-10) in October 2016. For each dataset, we have chosen a
couple of classes from different natures. We studied the completeness of resources
that have classes as the following ones: C = {Film, Organisation, Scientist,
PopulatedPlace}. For the properties used in the resources descriptions, we have
chosen the English datasets ”mapping-based properties”, ”instance types” and
”labels”.
The experiments revealed that datasets completeness could increase or de-
crease due to changes made to existing data or to the new added data. We also
noticed that often this evolution does not benefit 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 different 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 Conciseness assessment
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 undefined classes and properties [9], this leads to the lack
of semantic features in DBpedia ontology as illustrated in Table 4, only the prop-
erty domain-range restrictions can be applied. Unfortunately, only 30 functional
properties have been defined. Furthermore, DBpedia ontology neither defines
min and max cardinality nor the functional predicates nor transitive proper-
ties nor the symmetric ones. In addition, we noted that according to the last
version of DBpedia (October-2016), 16.3% of predicates are represented with-
out domains and 10.2% of predicates are without ranges in DBpedia ontology.
For this reason, based on the approach that has been proposed in [16], we infer
missed domains (and/or ranges) of predicates in DBpedia ontology. In case of
instances which have more than one rdf:type, only the class with the highest
value will be defined as the domain (or range) of the property if this value is
greater than a selected threshold. When the highest value is smaller than the
threshold, owl:Thing will be selected as the domain (or range). We applied our
approach on the last version of DBpedia ontology v2016-10. Table 5 shows top 10
results from the DBpedia dataset ranked by schema analysis. We chose support
thresholds 0.1% for the content filtering part.
Predicate 1 Predicate 2
Feature existence musicComposer composer
Domain 83.7% author writer
Range 89.8% creator author
Functional properties 1% starring artist
Transitive properties 0% city locationCity
symmetric properties 0% education almaMater
max cardinality 0% headquarter city
min cardinality 0% occupation personFunction
Table 4: Characteristics predicates of musicComposer musicalArtist
DBpedia dataset (v10-2016) musicBy musicComposer
Table 5: Top 10 matched predicate pairs
7 Approach
According to the research questions and the hypotheses formulated, we ad-
dress two aspects of linked data quality dimensions.
7.1 Completeness assessment
We represent, in this section, our approach [10] that addresses a completeness
aspect of linked data by posing the problem as an itemset mining problem. In
fact, the completeness at the data level assesses missing values [14]. This vision
requires a schema (e.g. a set of properties) that needs to be inferred from the
data source. However, it is not relevant to be considered for a subset of resources.
However, the schema is as the union of all properties used in their description
as seen in Section 1.1. Indeed, this vision neglects the fact that missing values
can express inapplicability.
Our mining-based approach includes two steps:
1. Properties mining: Given a dataset D, we first represent the properties,
used for the description of the D instances, as a transaction vector. We then
apply the well-known FP-growth algorithm [8] for mining frequent itemsets
(we chose FP-growth for efficiency reasons, any other itemset mining algo-
rithm could obviously be used). Only a subset of these frequent itemsets,
called ”Maximal” [6], is captured. This choice is motivated by the fact that,
on one hand, we are interested in important properties for a given class that
should appear often, and on the other hand, the number of frequent patterns
could be exponential when the transaction vector is very large.
2. Completeness calculation: Once the set of maximal frequent itemsets
MFP is generated, we use the apparition frequency of items (properties)
in MFP to give each of them a weight that reflects how important the set
of properties is considered for the description of instances. Weights are then
exploited to calculate the completeness of each transaction (regarding the
presence or absence of properties) and, hence, the completeness of the whole
dataset.
Definition 1. ( Completeness) Let I 0 a subset of instances, T the set of
transactions constructed from I 0 , and MFP a set of maximal frequent pat-
tern. The completeness of I 0 corresponds to the completeness of its transac-
tion vector T obtained by calculating the average of the completeness of T
regarding each pattern in MFP. Therefore, we define the completeness CP
of a subset of instance I 0 as follows:
|T | |MF P|
1 X X δ(E(tk ), P̂j )
CP(I 0 ) = (1)
|T | j=1
|MFP|
k=1
n
such that: P̂j ∈ MFP, and δ(E(tk ), P̂j ) = 1 if P̂j ⊂ E(tk )
0 otherwise
7.2 Conciseness assessment
The objective of the semantic analysis is to find the meaning of the pred-
icate. Depending on systematic analysis alone is not sufficient to discover the
synonymously used predicates, also too many false positive results are repre-
sented, 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 mean-
ing of each candidate. In addition, we study some conditions to examine them
exploring their meanings so that we mathematically prove on a basis of Descrip-
tion 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 predi-
cate 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 syn-
onym of other predicate in some cases according to the semantic features of each
one, such as: Disjoint properties based on their domains and ranges, Symmet-
ric/Asymmetric Property, Inverse Functional property, Functional property and
max cardinality. We illustrate these arguments using Description Logic formal-
ization.
8 Evaluation plan
Our goal is to compare the completeness and conciseness of dataset with our
presented approach to the state of the art such as Sieve [13]. In the future, we
plan to enrich our investigation with other data sources such as Yago, IMDB,
etc. In addition, for conciseness we would compare our approach to the Abedjan
approach [1] to judge the importance of the semantic analysis, we aim to prove
that the excluded candidates cannot be synonyms predicates. As the results show
that DBpedia dataset misses lots of metadata, we plan to find an approach to
infer the features of the predicates since we believe that we can help to improve
the use of the semantic part.
9 Reflections
Since we believe that poor-quality data affects 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 sufficiently 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 Cherfi and Dr. Fayçal Hamdi for their support and the opportunity for
the realization of this work.
References
1. Abedjan, Z., Naumann, F.: Synonym analysis for predicate expansion. In: Extended
Semantic Web Conference. pp. 140–154. Springer (2013)
2. Balaraman, V., Razniewski, S., Nutt, W.: Recoin: Relative completeness in wiki-
data. In: Companion of the The Web Conference 2018 on The Web Conference
2018. pp. 1787–1792. International World Wide Web Conferences Steering Com-
mittee (2018)
3. Batini, C., Cappiello, C., Francalanci, C., Maurino, A.: Methodologies for data
quality assessment and improvement. ACM Computing Surveys (CSUR) 41(3),
16 (2009)
4. Debattista, J., Auer, S., Lange, C.: Luzzu–a framework for linked data quality
assessment. In: Semantic Computing (ICSC), 2016 IEEE Tenth International Con-
ference on. pp. 124–131. IEEE (2016)
5. Fürber, C., Hepp, M.: Swiqa-a semantic web information quality assessment frame-
work. In: ECIS. vol. 15, p. 19 (2011)
6. Grahne, G., Zhu, J.: Efficiently using prefix-trees in mining frequent itemsets. In:
Proceedings of the ICDM 2003 Workshop on Frequent Itemset Mining Implemen-
tations, 19 December 2003, Melbourne, Florida, USA (2003)
7. Guéret, C., Groth, P., Stadler, C., Lehmann, J.: Assessing linked data mappings
using network measures. In: Extended Semantic Web Conference. pp. 87–102.
Springer (2012)
8. Han, J., Pei, J., Yin, Y., Mao, R.: Mining frequent patterns without candidate
generation: A frequent-pattern tree approach. Data Min. Knowl. Discov. 8(1), 53–
87 (Jan 2004)
9. Hogan, A., Harth, A., Passant, A., Decker, S., Polleres, A.: Weaving the pedantic
web. LDOW 628 (2010)
10. Issa, S., Paris, P., Hamdi, F.: Assessing the completeness evolution of dbpedia:
A case study. In: Advances in Conceptual Modeling - ER 2017 Workshops AHA,
MoBiD, MREBA, OntoCom, and QMMQ, Valencia, Spain, November 6-9, 2017,
Proceedings. pp. 238–247 (2017)
11. Joseph M., J., Richard S., B., Frank M., G.: The Quality Control Handbook.
Rainbow-Bridge, 3 edn. (1974)
12. Kontokostas, D., Westphal, P., Auer, S., Hellmann, S., Lehmann, J., Cornelissen,
R., Zaveri, A.: Test-driven evaluation of linked data quality. In: Proceedings of the
23rd international conference on World Wide Web. pp. 747–758. ACM (2014)
13. Mendes, P.N., Mühleisen, H., Bizer, C.: Sieve: linked data quality assessment and
fusion. In: Proceedings of the 2012 Joint EDBT/ICDT Workshops. pp. 116–123.
ACM (2012)
14. Pipino, L.L., Lee, Y.W., Wang, R.Y.: Data quality assessment. Communications
of the ACM 45(4), 211–218 (2002)
15. Ruckhaus, E., Baldizán, O., Vidal, M.E.: Analyzing linked data quality with li-
quate. In: OTM Confederated International Conferences” On the Move to Mean-
ingful Internet Systems”. pp. 629–638. Springer (2013)
16. Töpper, G., Knuth, M., Sack, H.: Dbpedia ontology enrichment for inconsistency
detection. In: Proceedings of the 8th International Conference on Semantic Sys-
tems. pp. 33–40. ACM (2012)
17. Zaveri, A., Rula, A., Maurino, A., Pietrobon, R., Lehmann, J., Auer, S.: Quality
assessment for linked data: A survey. Semantic Web 7(1), 63–93 (2016)