=Paper=
{{Paper
|id=None
|storemode=property
|title=Using Provenance for Quality Assessment and Repair in Linked Open Data
|pdfUrl=https://ceur-ws.org/Vol-890/paper2.pdf
|volume=Vol-890
|dblpUrl=https://dblp.org/rec/conf/semweb/FlourisRPMF12
}}
==Using Provenance for Quality Assessment and Repair in Linked Open Data==
https://ceur-ws.org/Vol-890/paper2.pdf
Using Provenance for Quality Assessment and Repair
in Linked Open Data
Giorgos Flouris1 , Yannis Roussakis1 , Marı́a Poveda-Villalón2 , Pablo N. Mendes3 , and
Irini Fundulaki1,4
1
FORTH-ICS, Greece, 2 UPM, Spain, 3 FUB, Germany, 4 CWI, The Netherlands
{fgeo,rousakis,fundul}@ics.forth.gr, mpoveda@fi.upm.es,
pablo.mendes@fu-berlin.de
Abstract. As the number of data sources publishing their data on the Web of
Data is growing, we are experiencing an immense growth of the Linked Open
Data cloud. The lack of control on the published sources, which could be untrust-
worthy or unreliable, along with their dynamic nature that often invalidates links
and causes conflicts or other discrepancies, could lead to poor quality data. In
order to judge data quality, a number of quality indicators have been proposed,
coupled with quality metrics that quantify the “quality level” of a dataset. In ad-
dition to the above, some approaches address how to improve the quality of the
datasets through a repair process that focuses on how to correct invalidities caused
by constraint violations by either removing or adding triples. In this paper we ar-
gue that provenance is a critical factor that should be taken into account during
repairs to ensure that the most reliable data is kept. Based on this idea, we propose
quality metrics that take into account provenance and evaluate their applicability
as repair guidelines in a particular data fusion setting.
1 Introduction
The Linked Open Data (LOD) cloud is experiencing rapid growth since its conception
in 2007. Hundreds of interlinked datasets compose a knowledge space which currently
consists of more than 31 billion RDF triples. In this setting, data constantly evolves,
invalidating previous links between datasets and causing quality problems; similarly,
changes in the world itself are not simultaneously reflected in all related datasets, caus-
ing conflicts and other discrepancies among overlapping datasets during data fusion.
Such data quality problems come in different flavors, including duplicate triples,
conflicting, inaccurate, untrustworthy or outdated information, inconsistencies, invalidi-
ties and others [23,1,14], and cost businesses several billions of dollars each year [7].
Therefore, improving the quality of datasets in an evolving LOD cloud is crucial.
Quality is generally defined as fitness for use [14]. Therefore, the interpretation of
the quality of some data item depends on who will use this information, and what is the
task for which they intend to employ it. While one user may consider the data quality
sufficient for a given task, it may not be sufficient for another task or another user. Thus,
it has been argued that the concept of quality is multi-dimensional, as well as context-
and application-specific [23]. To assess quality, a non-exhaustive list of quality dimen-
sions such as timeliness, trustworthiness, conciseness and validity has been considered
in [23,1,26]. A conceptual model for quality assessment is described, that is composed
of quality indicators, quality assessment metrics and scoring functions that quantify the
quality of a dataset along a given (set of) quality dimension(s) [26].
Apart from evaluating a dataset’s quality, it is also important to improve it, i.e.,
repair the dataset. We focus on evaluating automated repair methods [24,28] which rely
on a set of preferences that can be used as “guidelines” by the system to determine how
to resolve quality problems. For example, in the case of conflicting information, one
could opt to keep the most recent information to resolve the conflict (in accordance to
the Principle of Primacy of New Information often employed in evolution settings [17]).
The purpose of this paper is to determine how provenance can help in devising useful
preferences for improving the quality of LOD datasets.
Provenance refers to the origin or source of a piece of data and encodes from where
and how the piece of data was obtained [32]. Provenance is of paramount importance,
as in some cases it is considered more important than the data itself [8]. It is essential
in many applications, as it allows to effectively support trust mechanisms, digital rights
and privacy policies, and is also a means to assess the reliability of information; we
exploit the latter property to devise useful metrics for quality assessment and repair.
We focus on validity, which requires that the dataset conforms to a set of custom
constraints, expressed as logical rules. Validity is one of the most flexible and impor-
tant metrics, because it encodes context- and application-specific requirements. It has
been used in different contexts to express constraints like transitivity or functionality of
properties, cardinality constraints, foreign key constraints etc [3,18,25,30].
Previous work on repairing LOD datasets has focused on preferences related to
either the data itself [28], or to the metadata thereof (called fusion functions) [24]. This
paper combines the repairing algorithm of [28] and the approach advocated in [24] and
proposes the use of complex preferences that consider both data and metadata with an
emphasis on provenance to evaluate how this combination performs in a real setting
where data from disparate sources are fused to produce a LOD dataset. In a nutshell,
the main contributions of this paper are the following:
– The description of a set of provenance-based assessment metrics (Section 4).
– The extension of the repairing algorithm in [28] to support preferences on prove-
nance and other metadata (Section 5).
– The evaluation of the extended repairing algorithm under the proposed preferences
in a LOD fusion setting (Section 6).
2 Motivating Example
For illustration purposes, we consider a user who wants to find information about
Brazilian cities by fusing DBpedia dumps from different languages. Dumps overlap,
have different coverage, and not all of them contain correct or up-to-date information.
Therefore, the result of fusion may contain redundant, conflicting or inconsistent in-
formation (cf. Section 6), i.e., data of poor quality. To improve data quality, one could
remove conflicting information keeping one value for each city. Given that the user has
no access to some authoritative data source (if he did, then searching in the DBpedias
2
would be unnecessary), the only way to choose the correct value is to use heuristics, ex-
pressed as preferences, that determine the most reliable information. Such preferences
could involve the trustworthiness of information (based on provenance metadata), re-
cency, common sense, or some combination of the above metrics.
Doing so manually may be difficult or impossible, given that DBpedia contains 1400
conflicts of this type for Brazilian cities alone (cf. Section 6). Thus, it makes more sense
to use an automated repairing algorithm such as those proposed in [28,24] to perform
the repairing based on user preferences. Unfortunately, existing automated approaches
for repairing LOD datasets are restricted in the kind of preferences they support. In this
paper, we show how existing approaches can be extended and employed to perform this
task, and evaluate our methods using the above fused dataset on Brazilian cities.
3 Preliminaries
3.1 Quality Assessment
Data quality is a multi-dimensional issue, whose exact definition depends on the dataset’s
expected use [14]. To model these facts, a list of quality dimensions were collected
in [23,1,26] to capture the most common aspects of data quality, such as timeliness, ver-
ifiability, completeness, relevancy, validity etc. In addition, [23,26] proposed a generic
quality assessment methodology based on associating each dataset with a numerical
value that represents its quality along the given dimension(s). This framework is based
on quality indicators, scoring functions and quality metrics.
A quality indicator is an aspect of the considered dataset that indicates the suit-
ability of the data for some intended use. Indicators are very diverse and can capture
several types of information, including both the information to be assessed itself, and its
metadata. As an example, the last modification date can be used as a quality indicator
of “freshness” (related to the timeliness dimension). A scoring function is a numerical
assessment representing the suitability of the data for the intended use, as determined
by the quality indicator. Continuing the above example, a scoring function for freshness
could return the number of days between two specified dates. Quality assessment met-
rics (or metrics for short) are used to measure information quality. Essentially, metrics
combine information from various quality indicators and scoring functions to determine
a numerical value that represents the quality of the data.
3.2 Quality Repair
Our repairing approach is based on [28] where data violating one or more validity
rules, expressed in formal logic, are repaired in a manner that respects a set of user-
defined specifications expressed as preferences. Due to space considerations, we will
only briefly describe that approach here, and refer the reader to [28] for details.
The approach of [28] considers DED rules (Disjunctive Embedded Dependencies) [5],
which are rules expressed in first-order logic that can capture several types of con-
straints, including transitivity and functionality of properties, cardinality constraints etc.
In [28], a simple methodology for identifying DED violations and the possible ways to
3
resolve them is described. A violation can usually be resolved in several different man-
ners, which necessitates the use of preferences to determine the optimal repair solution.
Preferences are based on the idea of formal preference models employed by the
database community [11], and are declarative specifications allowing one to describe
the ideal solution in an abstract manner (e.g., “minimum number of schema changes”).
Such a specification can be used by the system to automatically determine the preferred
resolution by comparing each of the options against the “ideal” one. Formally, prefer-
ences are based on features which are functions that assess the repairing options under
some dimension relevant for the preference (e.g., “number of schema changes”); fea-
tures are then either minimized or maximized, and combined using operators to form
more complex preferences. Due to space limitations, the reader is referred to [28] for
more details on the formal specification of preferences; more examples of preferences
are given in Sections 5, 6.
It should be noted that this specification allows various complex preferences to be
defined, but their applicability is limited on repairing options because features are de-
fined upon repairing options; thus, this definition cannot support preferences taking into
account the repair result and/or metadata information. This is addressed in Section 5.
4 Quality Assessment Using Provenance
We describe three basic and two complex quality assessment metrics based on both
data and metadata information (e.g., provenance). The presented metrics are not the
only possible ones; instead, we focused on a particular set that is applicable to our
evaluation scenario. In the future, we plan to evaluate alternative scenarios/metrics.
More specifically, we use the provenance-related indicators of source reputation,
freshness and plausibility. The first relates the quality of data with the perceived trust-
worthiness of its source. The second associates the quality of data with its recency,
assuming that old information is more likely to be outdated and unreliable. The last is
data-related and is used to identify typos, outliers or other obviously incorrect values.
Each of these indicators defines a simple quality assessment metric, but can also be
combined to form more complicated ones; we describe two options below.
The first is based on the idea that freshness alone may not be an adequate metric;
depending on the application, a piece of data may be up-to-date for several years (e.g.,
total area of a country), for a few minutes or hours (e.g, temperature), or only for a few
seconds (e.g., location of a moving car). Thus, a refined metric could first determine
whether a certain piece of data is outdated (depending on the application); if so, then its
assessment would be based solely on freshness (assessing “how much” it is outdated); if
not, then source trustworthiness should be used. We call this metric weighted freshness.
The second is based on the idea that the reliability of a source sometimes depends
also on the data itself. In our example, we could assume that the Portuguese Wikipedia
is the closest to the domain (Brazilian cities) and is likely to be better for more esoteric
things (small cities), whereas the English one, being the largest and most edited, is
likely to be more reliable for larger cities [31]. Thus, a sophisticated metric could take
into account the data itself before determining source trustworthiness. In our example,
the perceived reliability of the Portuguese Wikipedia (compared to the English one)
4
should increase when the triple considered is related to a small city, and decrease for
large ones. We will call this metric conditional source trustworthiness.
5 Quality Repair Using Provenance
To evaluate the use of provenance-related metadata as a means to identify the prefer-
ences that determine the optimal repairing options during a repair process, we should
lift the limitations of [28] which defined the features of preferences to be applicable
only on repairing options. This does not present major challenges: all we need to do is
extend the features’ definition. The new features can then be seamlessly added in the
existing framework of [28] and used for improved quality repair. Here, we explain the
ramifications of this extension and compare the extended version with the original one.
The first extension to be considered allows features to be applied on the result of
the repair process rather than the repairing options. This way, the repairing process
selects how to resolve invalidities based on the dataset (repair) that these choices lead
to, instead of based on the choices themselves. For example, to model the preference
“I want the resulting class hierarchy to have minimum depth”, one should use a feature
that measures the depth of the class hierarchy; such a feature makes sense to be defined
on the repair, not the repairing options (which are of the form: “add/delete triple t”).
The results in [28] indicate that this extension does not change the expressive power
of features. Indeed, given the original dataset D and the repairing options, one can
compute the repair; similarly, given D and the repair, the repairing options that led to
this repair can be computed. Therefore, for any given D there is a one-to-one corre-
spondence between a repair and the repairing options that led to it. Thus, a preference
defined upon repairing options can always be equivalently rewritten into a preference
defined upon the repairs, and vice-versa [28].
Note that this correspondence is true for any given D; thus, given a preference on
repairs, and in order to define the equivalent one on repairing options, one should devise
a complicated expression that takes into account D. The same is true for the opposite
rewriting. As an example, the preference “I want the resulting class hierarchy to have
minimum depth” makes more sense to be expressed upon repairs, even though it can,
in theory, be expressed as a preference upon repairing options if the input data D (to
be repaired) is considered. On the other hand, the preference “I want minimum number
of schema changes” makes more sense to be expressed upon repairing options, even
though, again, a complicated equivalent preference upon repairs that takes into account
the original dataset D is possible. Thus, even though the two options are equivalent in
theory, they are complementary from the practical perspective.
The second extension allows features to consider metadata related to either the input
data, or the repair, or the repairing options. This is a clear and very important extension
to the original definition, because it allows important features of data to be considered,
such as provenance, trustworthiness, reliability, timeliness and others, thus increasing
the range and variety of preferences that a user can define.
5
6 Evaluation
Our experiments focus on evaluating the usability of the preferences described in Sec-
tion 4. To do so, we used the algorithm described in [28] to identify and repair incon-
sistent information found in a fused DBpedia dataset on Brazilian cities coming from
five different input datasets (Wikipedias). The result of the repair was compared against
an authoritative dataset (the Gold Standard) in order to evaluate its quality. Our results
show that all preferences perform relatively well in terms of the accuracy of the repairs
performed, and that complex preferences are not necessarily better than simple ones:
the simple preference based on source reputation performs very well in all cases.
6.1 Experimental Setting
Both the Gold Standard and the input datasets1 contain data about the area, popula-
tion and founding date of Brazilian cities (areaTotal, populationTotal and foundingDate
properties respectively).
The Input datasets contain data extracted from the English, French, German, Span-
ish, and Portuguese Wikipedia infoboxes. Each piece of data is associated with its
provenance (recording the Wikipedia version of the corresponding source article, us-
ing RDF named graphs) and last modification date (of the corresponding article).
The Gold Standard dataset (GS) was obtained from an independent and official
source, namely the Instituto Brasileiro de Geografia e Estatı́stica (IBGE)2 . We converted
the data to RDF, generating a LOD dataset accessible via a SPARQL endpoint3 .
The only requirement imposed for the validity of the fused dataset is that the prop-
erties areaTotal, populationTotal and foundingDate are functional, i.e., that each city
should have at most one value for its area size, population and founding date. This can
be expressed as a DED rule using the formal expression: ∀city, pop1, pop2 : (city,
populationT otal, pop1)∧ (city, populationT otal, pop2) → (pop1 = pop2). Similar
DED rules can be written for the other properties. Each of those rules was violated sev-
eral times, as several duplicate conflicting records for these properties appeared in the
various Wikipedias. When such a rule is violated, there are only two resolution options,
namely deleting population count pop1 or pop2. The evaluation process used the five
preferences described below, based on the quality metrics discussed in Section 4.
PREFER PT is based on the source reputation metric, that imposes a preference order
on the information found in different input datasets. In our case, we used the following
trust order for the Wikipedias: Portuguese, English, Spanish, German, French. In case
of conflicts, the most trustworthy information prevails. To implement this preference,
we can define a feature that gives a “trust rank” to each triple based on its source and,
in case of conflict, uses the preference to select the triple that has the highest rank.
PREFER RECENT is based on freshness and provides a bias towards recent informa-
tion: in case of conflict, the most recent information is kept. It can be modeled using a
feature that assigns to each triple a value indicating how long ago it was last edited.
1
All datasets can be found at: www.oeg-upm.net/files/mpoveda/ISWC2012Main
2
geoftp.ibge.gov.br
3
geo.linkeddata.es/brasil/sparql
6
PLAUSIBLE PT is based on the plausibility metric that considers the actual data and
is used to determine whether a piece of information is “irrational”. In particular, if the
population is less than 500, the area less than 300 km2 and the founding date earlier
than year 1500, the triple is considered “irrational” and will always be dropped in case
of conflict. If both conflicting triples are “rational”, then we resort to the PREFER PT
preference to choose one of the two.
WEIGHTED RECENT : this preference is based on the “weighted freshness” metric.
In case of conflict, it evaluates the last update date of the two conflicting triples: if they
are found to be close (less than 3 months apart) they are considered equally up-to-date,
so the preference PREFER PT is used to choose the one to keep; otherwise, the most
up-to-date information prevails (i.e., PREFER RECENT is used).
CONDITIONAL PT : this preference is based on the “conditional source trustworthi-
ness” metric. Intuitively, it states that for small cities (less than 500.000 citizens) the
Portuguese Wikipedia is more reliable so the PREFER PT preference is used; for larger
cities, we use PREFER EN, which is similar to PREFER PT except that the trustworthi-
ness order of the English and Portuguese Wikipedias is swapped.
For the repair process, we used the algorithm of [28] with the above validity rules
and preferences. The repair process initially searches in the dataset for rule violations,
and resolves each violation independently using one of the above preferences. Each
repair result was compared against the GS and its quality was evaluated under various
dimensions (conciseness, consistency, validity, completeness and accuracy).
6.2 Experimental Results
Our analysis considered five different quality dimensions, namely conciseness, con-
sistency, validity, completeness and accuracy [23]. These dimensions were evaluated
against the final result of the repair under the different preferences.
Conciseness requires that no duplicates appear in the result, and is guaranteed to
be perfect as our algorithm trims duplicate triples as part of the repair process. Consis-
tency is defined as the lack of conflicting triples, whereas validity is defined as the lack
of rule violations. These two notions coincide for the particular rules considered. Again,
our algorithm guarantees perfect consistency and validity, as it eliminates all conflicts.
Completeness is the coverage of information about the domain of interest that the fused
dataset exhibits. For the particular input, it has been measured to be 77,02%. Com-
pleteness is not affected by our process, because our algorithm only deletes conflicting
triples, and does not add or change triples.
The accuracy dimension is the most interesting of all; accuracy is defined as “being
as close as possible to the actual state of affairs”, where the “actual state of affairs”
in our case is taken by the GS. Accuracy is evaluated by comparing the output result-
ing from each preference against the GS. However, it is important here to distinguish
between “inherent” inaccuracy and inaccuracy caused by the repair process.
For example, the population of Aracati (per the GS) is 69159, whereas in the fused
dataset it has two population counts: 69159 and 69616. The repair process will drop one
of the two values to resolve the conflict; if it chooses to keep the correct value (69159)
then the accuracy of the dataset is improved, otherwise the repair process contributes
7
Table 1. Evaluation Results
populationTotal
Preference Good Bad Good+Bad Optimal Sub-optimal Approximate Total
PREFER PT 224 10 234 654 56 710 944
PREFER RECENT 208 26 234 580 130 710 944
RATIONALITY PREFER PT 224 10 234 663 47 710 944
WEIGHTED PREFER RECENT 221 13 234 641 69 710 944
WEIGHTED PREFER RECENT 220 14 234 642 68 710 944
areaTotal (raw)
Preference Good Bad Good+Bad Optimal Sub-optimal Approximate Total
PREFER PT 0 11 11 14 419 433 444
PREFER RECENT 2 9 11 87 346 433 444
RATIONALITY PREFER PT 0 11 11 9 424 433 444
WEIGHTED PREFER RECENT 1 10 11 25 408 433 444
WEIGHTED PREFER RECENT 0 11 11 32 401 433 444
areaTotal (modified)
Preference Good Bad Good+Bad Optimal Sub-optimal Approximate Total
PREFER PT 5 2 7 156 91 247 254
PREFER RECENT 5 2 7 151 96 247 254
RATIONALITY PREFER PT 5 2 7 155 92 247 254
WEIGHTED PREFER RECENT 4 3 7 152 95 247 254
WEIGHTED PREFER RECENT 5 2 7 150 97 247 254
foundingDate
Preference Good Bad Good+Bad Optimal Sub-optimal Approximate Total
PREFER PT 6 3 9 0 3 3 12
PREFER RECENT 6 3 9 0 3 3 12
RATIONALITY PREFER PT 6 3 9 0 3 3 12
WEIGHTED PREFER RECENT 6 3 9 0 3 3 12
WEIGHTED PREFER RECENT 6 3 9 0 3 3 12
to the inaccuracy of the data. The first case (i.e., when the algorithm keeps the correct
data) is called a good choice, whereas the second is called a bad choice.
On the other hand, the city Oiapoque has two conflicting population counts, namely
20226 and 20426 in the fused dataset; its actual population, per the GS, is neither 20226
nor 20426, but 20509. Thus, the inaccuracy is inherent in the data, and cannot be af-
fected by the repair process, regardless of the repairing choice made. In this case, the
algorithm’s choice only affects accuracy in terms of “closeness” to the actual value:
20426 is closer to the reality, and is therefore better in terms of accuracy. If the algo-
rithm chooses 20426, we say that we have an optimal approximation choice, otherwise
we have a sub-optimal approximation choice.
Finally, note that non-conflicting records that are inaccurate compared to the GS are
ignored because our algorithm does not deal with non-conflicting records. In the fol-
lowing, we present the number of good, bad, optimal and sub-optimal choices for each
of the preferences considered (Table 1) and also compare the accuracy of the dataset be-
8
Fig. 1. Comparing the Accuracy of the Input and the Repairs
fore and after the repair (Figure 1). For clarity, our analysis is split in the three important
properties of our dataset (populationTotal, areaTotal, foundingDate).
Our experiments showed that 944 cities contained a duplicate population entry. In
234 of these cases (24,8%), the correct value could be found in the input, and in most of
these cases the algorithm managed to select it correctly (208-224 times, depending on
the preference – cf. Table 1). The simple PREFER PT, as well as the PLAUSIBLE PT pref-
erences give the best results, indicating that the Portuguese Wikipedia is indeed the most
reliable when it comes to Brazilian cities’ population. Surprisingly, PREFER RECENT
performs poorly, despite the fact that population is a dynamic property where up-to-
date information is usually more reliable. In the remaining 710 cases where the actual
value could not be found in the input, PLAUSIBLE PT had the best performance again
with 663 optimal choices, followed by PREFER PT with 654. The worst behavior was
exhibited by PREFER RECENT, with only 580 optimal choices.
In the case of areaTotal, we note that the results were poor for all preferences. Upon
further investigation, we realized that the problem was partially caused by the format of
the area values. In particular, the area measurements in the Portuguese Wikipedia were
generally off the actual one by 3 orders of magnitude. This was most probably related
to the different use of the thousands separator (“,”) and the decimal places separator
(“.”) in the English and Portuguese Wikipedia. Unfortunately, this was not consistent
throughout the values, as some values in the Portuguese Wikipedia used the English
notation, whereas others used the Portuguese one.
To evaluate this hypothesis, we pre-processed all area values coming from the Por-
tuguese Wikipedia by multiplying them with 1000, and re-ran the experiment. The re-
sults of the modified input look much better as the number of total violations dropped
significantly (from 444 to 254), indicating that many of the original violations were
caused by said extraction problem. The results for the various preferences are relatively
good, the best performance being exhibited again by the PREFER PT and RATIONAL -
IZED PREFER PT preferences (with minor differences from the rest).
The results related to the foundingDate property are the least interesting of the three,
because, as can be seen in Table 1, all the preferences exhibited the same behavior.
In addition, we evaluated the accuracy of the dataset before and after the repairing
process, by determining how much the value of each property (populationTotal, areaTo-
tal and foundingDate) differs from the corresponding value in the GS (before and after
the repair), as a percentage of the value in the GS. The results are shown in Figure 1. The
9
orange line represents the total accuracy of the input (for all 3 properties), whereas the
other five lines represent the accuracy of the repair for each preference. The x axis rep-
resents the accuracy (0% indicates accurate values); the y axis represents the number of
triples that have the corresponding accuracy. Obviously, the values differ depending on
whether the raw or the modified areaTotal input was considered. The figure shows that
the repair process improves accuracy, and that all 5 preferences give similar accuracy.
7 Related Work
Several works on quality assessment have appeared in the literature (see [1] for a sur-
vey). In [2] users are allowed to express quality assessment policies to filter informa-
tion from the Web. In [16] customizable assessment processes are formalized through a
Web Quality Assessment model. In [10] SPARQL queries were used to identify various
quality problems, whereas in [22] an XML-based model is proposed and used by a web
service cooperation broker to select the best data from different services.
In the context of the Semantic Web, repairing addresses malformed values and
datatypes or accessibility and derefencability issues [12], entity matching and disam-
biguation [13], and resolving inconsistencies, incoherencies or invalidities (ontology
debugging) [29]. The latter type is the most relevant to our work.
Most works on ontology debugging consider some Description Logic as the un-
derlying language and address the problem of removing logical contradictions. In our
setting, we consider custom validity rules over RDF data. Moreover, most works focus
on the problem of identifying inconsistencies, rather than resolving them [9]. In the
cases where the active resolution of inconsistencies is supported, the resolution is usu-
ally done manually or semi-automatically [15], possibly with the help of an interactive
tool (e.g., ORE [19], PROMPT [27], or Chimaera [21]). To the best of our knowledge,
the only automated repairing approaches in the area appear in [24] and [28] regarding
data fusion and ontology repair respectively.
The work described in [24] considers repairing in the context of data fusion and is
very similar to ours. The authors propose a configuration-based approach where repair-
ing is based on a user-defined conflict resolution strategy (similar to a preference) that
uses quality metrics and fusion functions. However, such strategies can only consider
metadata information, and cannot be combined to form complicated strategies. More-
over, the approach of [24] is restricted to handling conflicting information, i.e., it cannot
support arbitrary validity rules (e.g., DEDs). Thus, the present paper can be seen as tak-
ing the best of both [24] and [28]: on the one hand, it allows more complicated validity
rules in the spirit of [28]; on the other it lifts the limitations of both [24] and [28] on
preferences, by allowing complicated preferences over both the data and metadata.
Most works record provenance by either associating triples with a named graph [4,33]
(as done here) or by extending an RDF triple to a quadruple where the fourth ele-
ment represents the triple’s provenance [6,20]. These works vary in the semantics of
the fourth element, which can be used to represent provenance, context, access control,
trust, or other metadata information. The work of [8] addressed the problem of record-
ing provenance in the presence of RDFS inference, where the provenance of an implicit
triple is defined as the “combination” of the provenance of the implying ones.
10
8 Conclusions
Quality is an important issue in modern datasets, especially in the context of LOD
where several dynamic and potentially unreliable sources are being interlinked, with no
central control or curation. Quality assessment allows the evaluation of the quality of
datasets, whereas quality repair aims at improving quality, with emphasis on the validity
dimension. Apart from its value as a stand-alone process, repairing is also an integral
process in various contexts, such as data integration/fusion [24] and evolution [17].
This paper deals with quality assessment and repair of LOD datasets. We proposed
a number of quality metrics and applied them as preferences in the repairing process
defined in [28]. Our approach is similar in spirit to [24], but extends [23,28,24] by
providing more sophisticated provenance-based assessment metrics and by combining
data and metadata information in the definition of complicated preferences that are used
as guidelines for repair.
Our focus was not on developing a general approach to the problem (which can be
found at [23,24,26,28]), but on evaluating the usefulness of the proposed metrics and
preferences. As the extensive literature on data quality has shown, the usefulness of
such metrics is application- and context-dependent, so we focused our evaluation on
a specific data fusion setting. This is, to our knowledge, the first work evaluating an
automated repair algorithm in a LOD fusion setting using provenance metadata.
In the future, we plan to consider and evaluate more diverse and/or complicated
quality assessment metrics (and their associated preferences) for both the quality as-
sessment and the quality repair problems in different settings. Moreover, we plan to
consider integrating the above automated method with some kind of interactive prefer-
ence elicitation interface as an aid for users to formulate complicated preferences.
Acknowledgments
This work was partially supported by the projects PlanetData and BabelData.
References
1. C. Bizer. Quality-Driven Information Filtering in the Context of Web-Based Information
Systems. PhD thesis, Freie Universitat Berlin, 2007.
2. C. Bizer and R. Cyganiak. Quality-driven information filtering using the wiqa policy frame-
work. Web Semantics, 7(1):1–10, January 2009.
3. A. Calı̀, G. Gottlob, and A. Pieris. Advanced processing for ontological queries. Proceedings
of VLDB Endowment, 3:554–565, 2010.
4. J. Carroll, C. Bizer, P. Hayes, and P. Stickler. Named graphs, Provenance and Trust. In
WWW, 2005.
5. A. Deutsch. Fol modeling of integrity constraints (dependencies). In Encyclopedia of
Database Systems. 2009.
6. E. Dumbill. Tracking Provenance of RDF Data. Technical report, ISO/IEC, 2003.
7. W. Fan. Dependencies revisited for improving data quality. In PODS-08, 2008.
8. G. Flouris, I. Fundulaki, P. Pediaditis, Y. Theoharis, and V. Christophides. Coloring rdf
triples to capture provenance. In ISWC-09, 2009.
11
9. G. Flouris, D. Manakanatas, H. Kondylakis, D. Plexousakis, and G. Antoniou. Ontology
change: Classification and survey. Knowledge Engineering Review (KER), 23(2), 2008.
10. C. Fürber and M. Hepp. Using semantic web resources for data quality management. In
EKAW-10, 2010.
11. P. Georgiadis, I. Kapantaidakis, V. Christophides, E. Mamadou Nguer, and N. Spyratos. Ef-
ficient rewriting algorithms for preference queries. In ICDE-08, 2008.
12. A. Hogan, A. Harth, A. Passant, S. Decker, and A. Polleres. Weaving the pedantic web. In
LDOW-10, 2010.
13. A. Hogan, A. Zimmermann, J. Umbrich, A. Polleres, and S. Decker. Scalable and distributed
methods for entity matching, consolidation and disambiguation over linked data corpora.
Journal of Web Semantics, 10, 2012.
14. J. Juran. The Quality Control Handbook. McGraw Hill, New York, 3rd edition, 1974.
15. A. Kalyanpur, B. Parsia, E. Sirin, and B. Cuenca Grau. Repairing unsatisfiable concepts in
owl ontologies. In ESWC-06, 2006.
16. T. Knap and I. Mlýnková. Web quality assessment model: trust in qa social networks. In 8th
international conference on Ubiquitous intelligence and computing, 2011.
17. G. Konstantinidis, G. Flouris, G. Antoniou, and V. Christophides. A formal approach for
rdf/s ontology evolution. In ECAI-08, 2008.
18. G. Lausen, M. Meier, and M. Schmidt. Sparqling constraints for rdf. In EDBT-08, 2008.
19. J. Lehmann and L. Buhmann. Ore - a tool for repairing and enriching knowledge bases. In
ISWC, 2010.
20. R. MacGregor and I.-Y. Ko. Representing Contextualized Data using Semantic Web Tools.
In Practical and Scalable Semantic Systems, 2003. In conjunction with ISWC.
21. D.L. McGuinness, R. Fikes, J. Rice, and S. Wilder. An environment for merging and testing
large ontologies. In KR, 2000.
22. M. Mecella, M. Scannapieco, A. Virgillito, R. Baldoni, T. Catarci, and C. Batini. Managing
data quality in cooperative information systems. In Meaningful Internet Systems, 2002.
23. P. Mendes, C. Bizer, Z. Miklos, J.-P. Calbimonte, A. Moraru, and G. Flouris. D2.1: Concep-
tual model and best practices for high-quality metadata publishing. PlanetData Del., 2012.
24. P. Mendes, H. Muhleisen, and C. Bizer. Sieve: Linked data quality assessment and fusion.
In LWDM-12, 2012.
25. B. Motik, I. Horrocks, and U. Sattler. Bridging the gap between owl and relational databases.
In WWW-07, 2007.
26. F. Naumann. Quality-driven query answering for integrated information systems. Springer-
Verlag, Berlin, Heidelberg, 2002.
27. N.F. Noy and M.A. Musen. Prompt: Algorithm and tool for automated ontology merging
and alignment. In AAAI, 2000.
28. Y. Roussakis, G. Flouris, and V. Christophides. Declarative repairing policies for curated
KBs. In Proceedings of the 10th Hellenic Data Management Symposium (HDMS-11), 2011.
29. S. Schlobach and R. Cornet. Non-standard reasoning services for the debugging of descrip-
tion logic terminologies. In IJCAI-03, 2003.
30. G. Serfiotis, I. Koffina, V. Christophides, and V. Tannen. Containment and minimization of
rdf(s) query patterns. In ISWC-05, 2005.
31. E. Tacchini, A. Schultz, and C. Bizer. Experiments with wikipedia cross-language data fu-
sion. In SFSW-09, 2009.
32. W.C. Tan. Provenance in databases: Past, current, and future. Bulletin of the IEEE Computer
Society Technical Committee on Data Engineering, 2007.
33. E. Watkins and D. Nicole. Named Graphs as a Mechanism for Reasoning About Provenance.
In Frontiers of WWW Research and Development - APWeb, 2006.
12