As Linked Open Data is gaining traction, publishers incorporate more their data to the cloud. Since the whole Web of Data cannot be semantically represented though, data consumers should also be able to map any content to rdf on-demand to answer complicated queries by integrating information from multiple heterogeneous sources distributed over the Web or not. In both cases, the quality and integrity of the generated rdf output a ects the performance of traversing and querying the Linked Open Data. Thus, well-considered and automated approaches to semantically represent and interlink, already during mapping, the domain level information of distributed heterogeneous sources is required. In this paper, we outline a plan to tackle this problem: We propose a uniform way of de ning how to map and interlink data from heterogeneous sources, alternative approaches to perform the mappings and methods to assess the quality and integrity of the resulting Linked Data sets.
E ciently extracting and integrating information from diverse, distributed and heterogeneous sources to enable rich knowledge generation that can more accurately answer complicated queries and lead to e ective decision making, remains one of the most signi cant challenges. Nowadays, Semantic Web enabled technologies become more mature and the rdf data model is gaining traction as a prominent solution for knowledge representation. However, only a limited amount of data is available as Linked Data, because, despite the signi cant number of existing tools, acquiring its rdf representation remains complicated.
Deploying the ve stars of the Linked Open Data schema1 is still the de-facto way of incorporating data to the Linked Open Data (lod) cloud. Approaching though the stars as a set of consecutive steps and applying them to separately individual sources, disregarding possible prior de nitions and links to other entities, leads in failing to reach the uppermost goal of publishing interlinked data. Manual alignment to their prior appearances is often performed by rede ning their semantic representations, while links to other entities are de ned after the 1 http://5stardata.info/ data is mapped and published. Identifying, interlinking or replicating, and keeping them aligned is complicated and the situation aggravates the more data is mapped and published. Existing solutions tend to generate multiple Unique Resource Identi ers (uris) for the same entities while duplicates can be found even within a publisher's own datasets. Hence, demand emerges for a well-considered policy regarding mapping and interlinking of data in the context of a certain knowledge domain, either to incorporate the semantically enriched data to the lod or to answer a query on-the- y.
So far, there is neither uniform mapping formalisation to de ne how to map and interlink heterogeneous distributed sources into rdf in an integrated and interoperable fashion nor complete solution that supports the whole mapping and interlinking procedure together. Apart from few domain speci c tools, none of the existing solution o er the option to automatically detect the described domain and propose corresponding mapping rules. Except for the eld of plain text analysis where again the main focus is on semantically annotating the text rather than describing a domain and the relationships between its entities. Moreover, there are are no means to validate and check the consistency, quality and integrity of the generated output, apart from manual user-driven controls, and no means to automate these tests and incorporate them in the mapping procedure. 2
The problem is directly relevant to data publishing and data consumption with an emphasis on semantically-enabled data integration. In the data publishing end of spectrum, domain level information can be integrated from a combination of heterogeneous sources and published as Linked Data, using the rdf data model. In the data consumption end of spectrum, the relevancy is two-fold: (i) On the one hand, the quality and integrity of the resulting rdf representation is re ected at the dataset' s consumption. (ii) On the other hand, data extracts can be mapped and interlinked on-demand and on-the- y from di erent heterogeneous sources, since not all data can be represented as Linked Data. On the whole, the problem is relevant to the alignment and synchronisation of data's semantic and non-semantic representations; modi cations (inserts, updates and deletions) need to be synchronised over data's semantic and non-semantic representation.
The problem is emphasized in cases of knowledge acquisition, searching or query answering that information integration is required from a combination of distributed and heterogeneous (semantic and/or non-semantic) data sources. Especially when it is taken into consideration data that cannot be easily traversed else, for instance the deep Web or large volumes of published data les. Semantic Web technologies together with the rdf data model allows to deliberately concatenate the extract of data that is relevant.
There are several stakeholders that could take advantage of such information integration enhanced with semantic annotation. Such key stakeholders are those who publish and consume large volumes of data that might be distributed and appear in heterogeneous formats. For instance, governments that publish and consume, at the same time, Open Data, scientists that combine data from di erent sources and re-publish processed information or (data) journalists that need extracts of data from several sources to acquire knowledge and draw conclusions. 3
Several solutions exist to execute mappings from di erent le structures and
serialisations to rdf. Di erent mapping languages beyond r2rml were de ned [
A larger variety of solutions exist to map data from xml to rdf, but to
the best of our knowledge, no speci c languages were de ned for this, apart
from the w3c standardized grddl5 that essentially provides the links to the
algorithms (typically represented in xslt) that maps the data to rdf. Instead,
tools mostly rely on existing xml solutions, such as xslt (e.g., Krextor [
In general, most of the existing tools deploy mappings from a certain source
format to rdf (per-source approaches ) and only few tools provide mappings
from di erent source formats to rdf. Datalift [
Beyond pure execution of mappings to rdf, most of the existing tools do not provide any recommendations regarding how the data should be mapped, 2 http://www.w3.org/2001/sw/rdb2rdf/wiki/Implementations 3 https://github.com/cygri/tarql 4 https://github.com/knudmoeller/Vertere-RDF 5 http://www.w3.org/TR/grddl/ 6 http://www.gac-grid.de/project-products/Software/XML2RDF.html 7 http://daverog.github.io/tripliser/ 8 http://thedatatank.com 9 http://www.isi.edu/integration/karma/ 10 http://openrefine.org/ 11 http://simile.mit.edu/wiki/RDFizers 12 http://virtuoso.openlinksw.com/dataspace/doc/dav/wiki/Main/VirtSponger namely how to model the domain described. Only Karma o ers mapping recommendation, however it relies on a training algorithm that improves when several domain-relevant data sources are mapped. Among the other tools, only Open Re ne supports recommendations to a certain extend, but its recommendations have the form of disambiguating named entities appearing in the lod.
As described, existing tools are solely focused on mapping data to the rdf model, rather than interlinking the entities of the source to existing entities appearing on the Web. Only Open Re ne allows to reconcile and match entities to resources published as Linked Data and Datalift which incorporates interlinking functionality but only as a subsequent step executed after the mapping is completed. Overall, till nowadays, mapping and interlinking are considered two steps that are executed consecutively. A lot of work has been done in the eld of text analysis, natural language processing (nlp) and named entity recognition (ner) to identify and disambiguate entities with resources appearing in the lod cloud. However such techniques are mainly focused on semantically annotating the text rather than modelling the domain described. Moreover these techniques are not applied in the case of (semi-)structured mappings.
Last but not least, none of the existing tools o er a complete solution that
allows to re ne the executed mappings based on the users' feedback, the results of
data cleansing tools, reasoning over the ontologies used or studying the integrity
and connectedness of the resulting dataset considering it as a graph. A summary
of existing approaches for assessing data quality that could be incorporated
for re ning the mappings according to the result of a mapping can be found
at [
{ How can we access and represent domain level information from distributed heterogeneous sources in an integrated and uniform way?
{ How can we enable querying distributed heterogeneous sources on the Web in a uniform way? On the other hand, the representation aspect needs to be investigated: { How can we identify if entities of a source have already been assigned a URI and enrich this unique representation with new properties and links? { How can we interlink newly generated resources with existing ones already during mapping considering the available domain information we have?
{ How can we assure that if we map some sources the domain is accurately modelled? { How well the entities of the dataset are linked with each other? { How well the dataset is linked with the lod cloud?
The main hypotheses related to my research are: { Integrated mapping and interlinking of data in heterogeneous sources generates fewer overlapping entities and models better the domain's semantics. { Reusing Unique Resource Identi ers (uris) leads to more robust and uniform datasets that have higher integrity and connectedness. { Interlinking such datasets raises the integrity and connectedness of the whole lod and improves the performance of its consumption. { Not all media can be published as Linked Open Data, thus mapping extracts of multiple heterogeneous data to rdf might occur on demand. 6
At this PhD, we propose a generic mapping methodology, that maps the data independently of the source structure (source-agnostic), puts the focus on mappings and their optimal reuse and considers interlinking already during mappings. Therefore, the initial learning costs remain limited, the potential for the custom-de ned mapping's reuse augments and a richer and more meaningful interlinking is achieved. This is a prominent advancement compared to the approaches followed so far. As a result, the per-source mapping model followed so far gets surpassed, leading to contingent data integration and interlinking. Beyond the language that facilitates the mapping rules' de nition and is the core of our solution, we propose a complete approach that aims to facilitate and improve the mappings de nition and execution.
In our proposed approach we aim to maximize the reuse of existing unique identi ers (uris) and rely on the links between them and the newly generated entities to achieve the interlinking of the new dataset with the lod. The disambiguated entities are assigned the corresponding uris and their representation is enriched with properties and relationships of the newly incorporated dataset. In contrast to the approaches followed so far, custom-generated uris are only assigned to the entities that were not identi ed in the lod cloud (not disambiguated). Based on the relationships between the newly generated entities and the disambiguated ones, the interlinking of the newly generated resources with the lod is achieved. In order to identify such entities, we propose applying ner techniques to the sources and use them against datasets of the lod.
Besides increasing the integrity of the dataset and reinforcing its interlinking with the lod cloud, the whole domain needs to be modelled. Recommendations based on vocabularies used for the description and for the relationships of the disambiguated entities or other entities that are identi ed to model the same domain and those appearing in a vocabularies' repository, such as lov13, can be taken into consideration. The domain can be further re ned after the execution of the mappings and the assessment of the output dataset using tools for evaluating 13 http://lov.okfn.org the data quality or taking into considerations the users' feedback. In these cases, the mapping rules can be adjusted to incorporate the emerging rules. 7
We already de ned a generic language adequate for de ning rules to map
heterogeneous sources into rdf in a uniform and integrated way [
Even though the language's extensibility is self-evident as rml relies on an
extension over r2rml, its scalability was also proven by further extending it to map
data published as html pages to the rdf data model. Results of the mappings
from html to rdf using rml were presented at the Semantic Web publishing
challenge of the 11th Extended Semantic Web Conference (ESWC14) [
A prototype processor15 was designed and implemented as a proof-of-concept
to accompany the rml mapping language. As rml extends r2rml, the
processor is implemented using an existing open-source r2rml processor16. The rml
processor was designed to have a modular architecture where the extraction and
mapping modules are independently executed and the extraction module can
be instantiated depending on the possible inputs. Short discussion regarding
alternative approaches for processors supporting rml were discussed at [
Finally, some preliminary work on mapping rules' re nements by
incorporating data consumers' feedback was presented at [
There are di erent aspects of the proposed solution which need to be assessed and we are aiming to evaluate: the RML mapping language itself, the semantic annotations and the entities interlinking, the quality and integrity of the resulting dataset and the performance of the mapping execution.
{ the language's potential in regard to (i) the range of input sources supported and their possible combinations for providing integrated mappings, namely 14 http://rml.io 15 https://github.com/mmlab/RMLProcessor 16 https://github.com/antidot/db2triples the language's scalability and extensibility ; (ii) the language's expressivity, namely the coverage of possible alternative mapping rules, mainly in comparison to other languages (or approaches) and (iii) last, how reusable and interoperable the mapping descriptions are. { the validity, consistency and relevance (especially when the domain is modelled according to automated recommendations) of the vocabularies used by the mapping rules to describe the domain knowledge. { the quality of the output. To achieve this, both automated solutions assessing data quality and domain experts will be used to evaluate the resulting dataset in regard to the identi ed or generated entities, the provided semantic annotations, the interlinking and the overall modelling of the domain. { the accuracy and the precision and recall of the retrieved, identi ed and enriched entities in conjunction with the con dence for the interlinked entities. { the integrity of the resulting dataset and the overall analysis of the output's datasets in respect to its graph-based representation, for instance in and out degree, its connectivity, its density, bridges, paths etc. { the impact of the resulting dataset's structure and interlinking in respect to its subsequent consumption. To be more precise, how traversing and querying the dataset is a ected by the choices taken while modelling the knowledge domain. In the case of querying, we aim to examine both the complexity of the queries de nition and the time and overload to execute them. { nally, while the performance is important to verify that the mappings can be executed in reasonable time, the performance of an rml processor is not the main focus of this work. However, the two fundamental ways of executing the mappings (mapping-driven or data-driven) will be evaluated and compared to identify best use-cases. The execution planning of the mapping rules though is more interesting and will be deeper investigated and evaluated. 9
The main di erence of our approach compared to existing works on mapping data is that we (i) introduce the idea of a uniform way of dealing with the mapping of heterogeneous sources and (ii) introduce the aspect of interlinking while we perform the mapping of data to the rdf data model. We approach the mapping from a domain modelling perspective where the data is either incorporated to a partially described domain or is mapped combined, forming their own domain. This way, we achieve generating datasets with higher integrity that are already interlinked among each other and with the lod and thus we reduce the e ort for subsequent interlinking of resources and o er better conditions for their subsequent consumption.
The research described in this paper is funded by Ghent University, the Flemish Department of Economy, Science and Innovation (EWI), the Institute for the Promotion of Innovation by Science and Technology in Flanders (IWT), the Fund for Scienti c Research-Flanders (FWO-Flanders), and the European Union.