Data linking is understood as the task of establishing typed links between entities across di erent RDF datasets via the help of automatic link discovery systems. Based on decades of research and practice in the Web community, the current paper speculates on the need of a paradigm shift when it comes to designing such systems. We depart from the premise that the current state-of-the-art focuses on genericness and automaticity, while not paying su cient attention to the particular properties and nature of the underlying data and the ensuing linking problem types. We draw new research axes, upon which the data linking task will be rede ned as automatic detection of the type of linking problem at hand based on the characteristics (pro les) of the candidate datasets.1;2 Linked data and its underlying technologies have been gaining popularity over the past years, due to the means they o er for data reuse and federation, increased visibility and sharing on the web and facilitated exchange of metadata. We de ne the problem of data linking as that of automatically establishing typed links between the entities of two or more RDF datasets or graphs. A variety of data linking systems have been proposed over the past 15 years within the Web community with interactions with government, cultural or research institutions as major linked data consumers and providers. As a result, vast amounts of linked data already exist on the Web (we refer, for example, to the LOD project). A number of benchmarks are developed and shared publicly in order to provide frameworks for the evaluation of data linking systems, driven by the well-known OAEI campaign, or the more industry-oriented EU HOBBIT project.3 Where are we now. State-of-the-art research into data linking [1] goes in two main directions: (1) proposing novel generic data linking systems and (2) developing methods for automatic link speci cation by (semi-)supervised machine learning techniques, in order to assist the con guration and tuning of established tools. Several of the most common systems, such as SILK [2] and LIMES [3], adopt a property-based link-discovery strategy: a set of predicates has to be 1 Copyright c 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). 2 The author thanks Stefan Dietze and Mathieu D'Aquin for their valuable feedback. 3 http://oaei.ontologymatching.org, https://project-hobbit.eu/
selected before the system proceeds to compare their values by the help of (an
aggregation of) similarity measures that also need to be selected and tuned. This
con guration task can be demanding in terms of user involvement in real-world
scenarios. Hence, a number of methods have been proposed to assist the users
in the con guration process. Properties to compare can be selected by the help
of key discovery tools. While many approaches exist (e.g., [
Most existing linking approaches have in common the fact that they attempt
to solve the problem from a generic stance by remaining vastly agnostic to the
nature of the underlying data [
We formulate the arguably outrageous hypothesis that there exist a nite
number of identi able and generalisable types of linking problems, de ned as
heterogeneity types that two to-be-linked datasets can manifest, e.g. di erences
in terminology, natural languages or structure, as presented in [
Proposed solution and challenges. Based on the hypotheses formulated above, we propose to direct future research and engineering e ort into the development of a data-centric bottom-up linking framework that channelizes and consolidates existing disparate e orts. This will allow to build on the wealth of linked RDF graphs via their in-depth analysis and consolidation and take advantage of years of research and practice in the eld (cf. Fig. 1). We identify a set of research axes and associated challenges on the way to realize this framework.
(1) Linked data harvesting and analysis. This axis consists in the consolidation of a large amount of already existing quality linked Web data, benchmarks, evaluation campaigns and linked data projects from a large variety of domains. An in-depth statistical analysis of these data will allow to identify a number of limited and generalisable linking problem types, discover correlations between application domains and data structure or quality, or between heterogeneity types and link density. This will inform the feature design in (2) and will generate training data for the automatic classi cation of pairs of datasets according to their linking problem type(s). Challenge: We need to ensure high quality of the links from which we will learn. Hence, we propose to rely on existing real-world benchmarks as a starting point, where datasets are often grouped according to speci c heterogeneity criteria (terminology, logics, structure, etc) that can be mapped to the identi ed linking problem types. Alternatively, one can apply existing linking methods of high precision in a preprocessing step. Relying on a large variety of data sets and domains is important to guarantee representativity.
(2) Joint datasets feature design. This axis involves data linking-oriented
graph pro ling and feature extraction via state-of-the-art RDF graph pro ling
techniques [
(3) Learning and applying ML models. This axis will rely on the training data harvested in (1) and the features extracted in (2) in order to de ne and apply classi cation models for linking problem type detection. Challenge: We identify here the standard challenge of selection and tuning of ML model(s) from a set of supervised algorithms. In addition, the multitude of possible classes will lead us beyond the standard binary classi cation task.
(4) Filling in \a shelf" of automatic, adaptable, modular solutions for each
of the linking problems identi ed in (1). We rely on the premise that a linking
problem type is ne-grained enough so that a particular modular solution can be
applied to it (for example, relying on lexical synset intersection for
synonymytype heterogeneity). These modular solutions will be identi ed by a comparative
analysis of the modules and respective performances of a large spectrum of
existing data linking systems, as this has been in part performed in [
Conclusion. Instead of trying to t a generic solution to any linking problem and dataset type, we suggest to enable a better understanding of the underlying data before applying a targeted solution best suited to the particular datasets at hand. We rely on the premise that the in-depth analysis of large amounts of linked data will allow to isolate a limited number of identi able data linking problems that ML models based on datasets pro les will help detect automatically. The moment is appropriate to take this approach for reasons of, one the one hand, the large and growing availability of linked data in an ever greater number of domains and, on the other hand, the existence of a large plethora of data linking tools, result of decades of research and practice. We hypothesise that channelizing these decentralised endeavours will foster and facilitate the application of linked data technologies within and across an even larger variety of domains and will ultimately free the domain expert of the technological burden.