One of the core functionalities of Knowledge Graphs is that data is not required to adhere to strictly de ned data models. Nonetheless, the RDF model provides speci cations for publishing data on the Web that not only describe entities and relationships but also focus on enabling interoperability between datasets. However, nding the right ontologies to model a dataset is a challenge since several valid data models exist and there is no clear agreement between them. We present a demonstration of an interface that allows users to customise a data model based on recommendations obtained with the RICDaM framework. This framework focuses on ranking candidates based on three metrics that measure the relevancy of the candidate, the interoperability of the neighbourhood of the candidate, and the overall consistency of the model proposed. The interface allows the user to re ne the recommended data model and complete it when the framework is lacking or not directly tting the intention of the user. The demo can be found at http://afel.insight-centre.org/ricdam/
Data models have an important role in data integration on the Web because they
de ne how data is connected and stored. The RDF data model uses
subjectpredicate-object statements (i.e., triples) to model datasets. The data model is
usually supported by one or more ontologies that provide standard nomenclature
to describe concepts. From the ontologies, classes are used to annotate the entity
types of subjects and objects in an RDF dataset, while datatype and object
properties are used to model predicates. A survey conducted over several years with
Linked Data providers found that the third most common barrier to
publishing Linked Data was Selecting appropriate ontologies to represent our data [
We propose the RICDaM framework (Recommending Interoperable and Consistent Data Models) which produces a ranked list of candidate data models that not only t the data but are also interoperable with a Knowledge Graph of multiple published RDF data sources. The output of the framework is the correspondence between a list of triple patterns (i.e., domain, property, range triple) from an input dataset and a ranked list of candidate triple patterns. We exploit the content and graph structure of this Knowledge Graph to compute scores that consider the accuracy, interoperability, and consistency of the candidates. These scores are combined into a single score that is weighted according to the user's preferences or use-case. In this demo, we present an interface that explores the use-case of aligning two datasets in the library data domain with published RDF library data models. The demonstrator shows the output of the framework and allows the user to choose the weights of the scores and manually customise the automatic recommendations generated by the framework.
The problem of creating a data model for a dataset is related to schema
matching, which maps relationships and concepts. A variety of schema matching
solutions have been proposed for di erent types of data (e.g., [
The RICDaM framework, illustrated in Figure 1, has three stages: (1) Building Background Knowledge, (2) Candidate Generation, and (3) Candidate Ranking. 2.1
Building Background Knowledge The goal of this stage is to build the background knowledge structures that will facilitate the next stages of the framework. The rst task includes creating a single KG from multiple existing RDF datasets. The entities of the KG are collected in a document store and indexed using an inverted index to enable full-text search. These entities maintain their original predicates but are also connected to external entities via the underlying ontology graph that facilitates candidate ranking. We enrich this ontology graph by inferring missing information (e.g., entity type of owl:SameAs relationships) and via ontology matching.
Knowledge Building
Data
Documents Entity Types
Document Store Entity Types Unranked Entity Type Candidates
Knowledge
Graph
DDaatataLLaayyeerr Data Layer
PPrPororpopeperetritreiteisess RDF Sources Pre-computed
metadata OOOnntnototloololgogygyyLLaLayayeyererr PrOopbejerctites PDraotpaetyrtpiees
Random Forest
Properties
Input Dataset (CSV, JSON, RDF)
Object and Datatype
Properties Unranked Object Unranked Datatype
Property Property
Candidates Candidates
Candidate Generation Content Score
Interoperability
Score
Consistency
Score Ranked Entity Type
Candidates
Ranked Data Model
Candidates Candidate Ranking
Content Score
Interoperability
Score Ranked Property
Candidates
In a nal step, we extract metadata from the document store (e.g., entity type frequency), and we train a Random Forest model on datatype property values of the KG to facilitate the generation of datatype property candidates. This stage generates a list of entity type and property candidates for an input dataset. For the entity types, the candidate list is obtained by searching literals in the inverted index of the document store. Datatype properties are obtained with the Random Forest classi cation model, which produces datatype property predictions for input literal values. Object properties are inferred from the relationships between the entity type candidates in the ontology graph.
We propose to rank candidates using a Content score and an Interoperability Score. The Content Score combines metrics based on string similarity, search results frequency, and graph distances. This score assesses the appropriateness of a candidate to match an entity type or property from an input dataset The Interoperability score contains information from the document store and from a sub-graph that restricts edges to relationships of equivalence, subsumption, or relatedness extracted via ontology matching techniques. This score measures how interoperable a candidate is considering the frequency of a candidate in the KG and the graph neighbourhood of the candidate. Higher Interoperability scores translate into a more connected and relevant candidate. This score uses the interoperability metric and the neighbourhood size. The neighbourhood size represents the total number of related neighbours up to a distance in the graph, while interoperability metric counts the frequency of these neighbours in the document store. In contrast with the interoperability metric, the neighbourhood size rewards neighbours that do not appear in the RDF data sources but are still connected to the candidate in the ontology graph.
Until now, candidates are ranked independently of each other. Therefore, the nal metrics compute the Consistency score. This score increases the likelihood of the same candidate being suggested for the same input entity type or property and, at the same time, boosts triples that are more commonly encountered in the Knowledge Graph. This consistency is achieved by ensuring that triples that appear together (co-occur) in the KG are rewarded, while also guarantees that the same triple elements are assigned the same entity types or properties throughout the data model. Therefore, the consistency combines Content, Interoperability, and co-occurrence frequencies into a single score that ranks candidate data model triples in terms of their adequacy and interoperability. 3
This demonstration presents the top candidate data model proposed to the user in response to an input dataset in the library domain. The interface gives an overview of the best ranked candidates for each triple but also allows the user to adapt the data model to their preference and use-case.
The main functionalities included in the demo are (1) an overview of the output of the framework, (2) customisation of the data model, (3) tuning of the parameters to produce di erent candidate rankings, and (4) exporting the data model as a set of mappings between the input and the produced data model.
When the user makes a manual change to the data model, they can choose to propagate that change to maintain consistency across the dataset or keep the change locally to the modi ed cell. The tuning parameters allow the user to customise the ranking of the candidates, obtaining di erent top data models that can speed up the modelling process by suggesting the candidates the user is looking for more easily. Finally, the user can export the data model produced and apply the mappings to translate their original input data to an RDF dataset that is potentially more interoperable with the datasets in the KG.
Table 1 presents the datasets used in the demo. The dashed line separates datasets used to build the background KG (top) and datasets used as input (bottom). Overall, the datasets contain a variety of records such as books, audio records, and periodicals. For example, for pgterms:ebook in Project Gutenberg, the suggestion is to use bibo:Document as the most interoperable, relevant, and consistent entity type, together with properties such as dcterms:contributor. Through the interface, it can be changed, for example, to schema:Book, which is often used as well in the background KGs.
https://www.bnf.fr Biblioteca Nacional de Portugal Portuguese http://www.bnportugal.gov.pt
http://www.bne.es
Gutenberg https://www.gutenberg.org https://openlibrary.org # Entities
In general, designing a data model is not a trivial task. When considering integration with existing datasets, the task becomes more complex. Our demo facilitates the task of nding the best possible data model according to certain criteria by producing a ranked list of candidates to match entity types and properties in a dataset. In the future, a complete tool using this framework would allow the user to search the indexed ontologies or add new ontologies to the graph to complete the model when the data model fails to nd the desired class. For JSON or CSV input datasets, it would also be possible to generate a RML mapping le to facilitate the process of translating the data to RDF.