Semantic Data Integration and the Linked Data Lifecycle

Nikola Tulechki, Vladimir Alexiev - Ontotext


0.1 Introduction

  • Semantic Data Integration is a lot more than Ontology Engineering:
    • Dataset research
    • Data analysis
    • Data cleaning
    • Semantic model: examples, shapes, documentation
    • Ontology engineering
    • Data mapping/conversion (ETL)
    • Instance matching/reconciliation
    • Data fusion/harmonization
    • Semantic text/metadata enrichment
    • Data enrichment, inference, aggregation
    • Sample queries
    • Semantic search, apps, visualizations

0.2 How Does This Differ from Data Warehousing?

  • Semantic databases (Knowledge Bases or Knowledge Graphs) are self-describing
  • Data warehousing usually focuses on statistical data (OLAP)
  • KGs can represent OLAP using the W3C Cube ontology, but can also represent any other kind of data
  • There are vast LOD datasets that can be utilized in building KGs
  • Everyone seems to be building a KG!

0.3 Example

Google KG: Entity Pages, Disambiguation

0.4 Competence Questions

  • Before developing applications, one better develop business requirements
  • Before procuring/integrating data, one better develop competence questions
  • These are sample questions that the KG should be able to answer
  • They should lead both dataset research and semantic modeling

1 Dataset Research

  • Given a data goal (competence questions), what datasets are available and relevant?
  • What is the enterprise data that we “intuitively” know should be captured?
  • Are there data gaps compared to the competence questions? How can we fill them?

1.1 How to Find Datasets?

1.2 LOD Cloud

The Linked Open Data (LOD) Cloud contained 28 datasets in 2007

1.3 … 89 datasets in 2009 …

1.4 … 295 datasets in 2011 …

1.5 … 570 datasets in 2014 …

1.6 … 1234 datasets today!

16136 links between datasets, 30B triples

1.7 Wikidata

Number of Creative Works and Cultural Institutions in Wikidata:

1.8 Finding Data

  • By Example:
    • Eg to find computer science awards, get a famous computer scientist (eg Tim Berners-Lee) and explore his awards
    • Eg to find artist biographic info, get a well-known artist (eg Emily Carr), find mentions of her on the web, list the ones that provide good biographic text
  • Knowledge and experience helps

2 Dataset Analysis

2.1 Scope

  • Feasibility estimations confronting the data with the use cases.
  • Familiarisation with the data
    • Size and complexity estimations of the data are performed
    • The data is comprehensively understood
    • Exceptional values are extracted and analyzed
  • Cardinalities are extracted and analyzed
  • The tools for the later phases are chosen
  • General feasibility is estimated
  • The groundwork for the mapping component of ETL is laid down

2.2 Key Value Analysis

  • Key values that often drive the mapping. Key-values could be mapped to:
    • Individual (eg skos:Concept)
    • Class
    • Property
    • or can even dictate a mapping decision, eg one mapping branch for Persons, another for Organizations

2.3 Key Value Example

Frоm Getty LOD - “Excel-driven Ontology Generation™”


2.4 Data Coverage

Example: PubMed Field Statistics

2.5 Data Quality

  • Coverage assessment dictates what is worth converting
  • Quality assessment drives data cleaning needs and decisions

3 Data Modeling

3.1 Scope

  • Create a data model for the domain of interest
  • Find existing relevant ontologies to use
  • Add custom classes and properties as needed (Ontology Engineering)
  • Document the model comprehensively: describe modeling patterns, justify decisions
  • Document in machine-readable way using RDF Shapes (SHACL or ShEx)
  • Make examples that can drive (generate) other artifacts, eg example diagrams, shapes

3.2 URL Design

3.3 Existing or New Ontology?


  • Reuse as much as possible
  • This will save you time, and will make your data more easily reusable
  • Search for ontologies: Linked Open Vocabularies


  • Multiple namespaces make data production/consumption a bit harder
  • is the ultimate “chauvinistic” example: single namespace, and even extensions (eg GoodRelations or SchemaBibEx) land in the same namespace.
    • I guess webmasters like the simplicity (and still often get it wrong ;-)
  • Don’t reuse a complex ontology for a single term (heavy ontology baggage)
  • Consider reusing ontology terms, but not necessarily loading the ontology
    • Example: every dct: property is a subprop of dc:, and such inference may be useless to you

3.4 Ontology Methodologies

How to do ontology engineering?

  • Avoid it if you can (i.e. reuse ;-)
  • Competence Questions !
  • Methods such as DILIGENT, METHONTOLOGY, NeON Methodology, ROO Kanga, SAMOD (Simplified Agile Methodology), HCOME (Human-centered ontology method), Aspect OntoMaven (Aspect-Oriented Ontology Development), IDEF5, ONTOCOM (cost estimation)
  • Ontology Design Patterns: typical situations. Composable!
  • Top-level ontologies: BFO, CCO, DOL/DOLCE, SUMO, UFO, Proton…
    • For cultural heritage: CIDOC CRM, ConML/CHARM

4 Data Conversion (ETL)

4.1 Scope

  • Transformation and homogenization of data sources into the target format, in order to populate the semantic model with instances
  • Important elements:
    • Choice of tools in accordance with the need and requirements
    • Maintainability of the solution
    • Reproducibility / exportability / portability
    • Data cleaning
  • Other Considerations:
    • formats
    • size scale (tools parse in memory)
    • consistency with existing conventions (project, sector)

4.2 Example

Onto ETL tools evaluation

5 Validation

5.1 RDF Shapes

TODO links

  • ShEx: W3C community spec. Pros: much briefer (compact, JSON and RDF representations), allows recursive data models, flexible focus nodes (shape map).
  • SHACL: W3C standard. SHACL-core and SHACL-standard. Advanced Features is a community spec. Pro: standardizes validation results
  • Validating RDF book (reviewed by Ontotext)
  • Implementations at Validating RDF wiki): summarized at next slides, but see the link for more details!

5.2 ShEx Implementations

name language playground, source, distribution
shex.js js,
ShEx NPM js
ShEx-validator js
Validata js,,
ShExJava java,,
RDFShape, ShaclEx scala,,,
TrucHLe scala
PyShEx python
shex.rb ruby
ShExkell haskell

5.3 SHACL Implementations

name language playground, source,distribution
SHACL rdf4 java
SHACL batch java
ELI-validator java,,
OSLO Validator java,,
shacl-runner scala
Netage SHACL java
SHACL-Check js
RDFShape, ShaclEx scala,,,
pySHACL python,
RDFUnit java
alt SHACL python

5.4 Custom Test Suites

  • RDFUnit (source, demo): sources custom patterns, OWL, OCLS shapes, DC Application Profiles, SHACL
  • Only SPARQL queries
    • negative examples
  • SPARQL queries and example output
    • Compares output of a query to the desired output
    • Could test data and/or the queries themselves
  • Domain Specific-Validation

6 Text and data enrichment

6.1 Scope

Semantic enrichment is the adding of value to a dataset by increasing the amount of queryable information it contains and/or decreasing the data’s noisiness. This is done in several ways:

  • Inference and link discovery
  • Thesaurus harmonisation
  • Entity mining
  • Instance matching and deduplication.
  • Data fusion

6.3 Thesaurus harmonisation

6.4 Entity mining.

  • Unstructured data (text, images) is processed in order to extract novel entities and relations and add them to the dataset.
  • Content classification
    • Attribution of categories to text, images or sound clips
    • Statistical methods
    • e.g e-mail filters
  • NLP Named entity recognition (NER)
  • Relation extraction from text
    • NER + relations between entities

6.5 Data fusion.

Redundant data is deduplicated and fused to produce a single master dataset without conflicts. selection of representative single fields (eg logo), * accumulation of multiple fields (eg names, transactions), * aggregation of summary fields (eg count or total amount)

6.6 Matching

  • Critical problem in data cleaning and integration
  • Overlapping instances across multiple datasets (client only or client and LOD) are matched and the sum of their attributes across datasets become available for querying.
  • Entity matching (EM) finds data instances that refer to the same real-world entity
  • We focus on EM as a process of transforming a string to a thing based on the provided semantic context

6.7 Basic Reconciliation

  • Fuzzy name matching
  • Simple additional features (e.g. exact string matches, differences in numbers)
  • Out-of-the-box match scoring & available recon services
  • Custom field parsing and normalization - additional parsing rules (acronyms, titles, dates etc)
  • Complex additional features
    • text analysis
    • hierarchical features
    • geographical features
    • network topology
  • Custom match scoring (possibly deep learning) e.g deep siamese text similarity

7 Modelling update flows

8 Model documentation

8.1 Scope

  • Data Diagrams
  • Detailed description of both new ontology terms (classes and properties) and reused ones (describing the specific use in our application profile).
  • Reference documentation
  • Semantic publishing of model
  • Sample Queries

8.2 Sample queries

  • Very handy way to augment the documentation
    • Can be used by a new user to get a feel of the data
    • Highly informative when combined with a short description
    • Stem out of competence questions
    • Can be used for testing purposes
  • Eg GVP sample queries

9 Semantic search apps and visualizations

9.2 Visualizations