1.
Knowledge Graph Maintenance
Prof. Paul Groth | @pgroth | pgroth.com | indelab.org
Thanks to Daniel Daza, Corey Harper, Thiviyan Thanapalsingam, Niels ten
Oever, Marieke van Erp, Valentin Vogelman and Frank van Harmelen
NEC Lab Europae
January 22, 2021

2.
We investigate intelligent systems that support
people in their work with data and information from
diverse sources.
In this area, we perform applied and fundamental
research informed by empirical insights into data
science practice.
Current topics:
• Automated Knowledge Base Construction
• Data Search + Data Provenance
• Data Management for Machine Learning
• Causality for machine learning on messy data
indelab.org

3.
Roads
and Bridges:
The Unseen Labor Behind
Our Digital Infrastructure
W R I T T E N B Y
Nadia Eghbal
Source:
https://www.fordfoundation.org/work/learning/research-reports/roads-and-bridges-the-unseen-labor-behind-our-
digital-infrastructure/

4.
Faculty of Science
Knowledge Graphs

5.
Source:
Azzaoui, K., Jacoby, E., Senger, S., Rodríguez, E. C., Loza, M., Zdrazil, B., … Ecker, G. F. (2013). Scienti
fi
c
competency questions as the basis for semantically enriched open pharmacological space development. Drug
Discovery Today, 18(17–18), 843–852. https://doi.org/10.1016/j.drudis.2013.05.008

6.
Source:
https://www.biocuration2019.org/about

7.
Source:
https://www.wired.com/story/inside-the-alexa-friendly-world-of-wikidata/

8.
Source:
https://stats.wikimedia.org/v2/#/en.wikipedia.org/contributing/user-edits/normal||2001-01-01~2019-09-01|~total|

9.
Crowdsourcing
100,000s of hand annotated examples
The TAC Relation Extraction Dataset
Source:
Zhang, Yuhao, et al. "Position-aware attention and supervised data improve slot
fi
lling." Proceedings of the 2017
Conference on Empirical Methods in Natural Language Processing. 2017.
Karen Fort, Gilles Adda, Kevin Bretonnel Cohen. Amazon Mechanical Turk: Gold Mine or Coal Mine?. Computational
Linguistics, Massachusetts Institute of Technology Press (MIT Press), 2011, pp.413-420. 10.1162/COLI_a_00057

10.
Concept1
Concept2 Concept3
KOS
Professional
Curators
Literature
Software
Non-professional
contributors
1. dealing with changing cultural and societal
norms, specifically to address or correct bias;
2. political influence
3. new concepts and terminology arising from
discoveries or change in perspective within a
technical/scientific community
4. gardening
5. incremental contributorship
6. progressive formalization
7. software and automation
8. integration of large numbers of data sources
9. variance in algorithm training data
Data
⚐
Society & Politics
(4, 5, 6)
(7, 8, 9)
(3)
(1, 2)
Source:
Michael Lauruhn and Paul Groth.
“Sources of Change for Modern Knowledge Organization Systems." Knowledge Organization 43, no. 8 (2016).

11.
Content
Universal
schema
Surface form
relations
Structured
relations
Factorization
model
Matrix
Construction
Open
Information
Extraction
Entity
Resolution
Matrix
Factorization
Knowledge
graph
Curation
Predicted
relations
Matrix
Completion
Taxonomy
Triple
Extraction
Concept
Resolution
14M
SD articles
475 M
triples
3.3 million
relations
49 M
relations
~15k ->
1M
entries
Paul Groth, Sujit Pal, Darin McBeath, Brad Allen, Ron Daniel
“Applying Universal Schemas for Domain Specific Ontology Expansion”
5th Workshop on Automated Knowledge Base Construction (AKBC) 2016
Knowledge Graph Curation and Re
fi
nement

12.
Apply ML

13.
Link Prediction

14.
Inductive Prediction
Daniel Daza, Michael Cochez, and Paul Groth. "Inductive Entity
Representations from Text via Link Prediction." arXiv preprint
arXiv:2010.03496 (2020). Accepted to WebConf 2021

15.
Inductive Prediction

16.
Inductive Prediction

17.
Inductive Prediction
Transformer

18.
Evaluation Scenarios
• Transductive (standard) - predict links between entities seen at
training time
• Dynamic - predict links between entities where a unseen entity at
training time can be in the head, tail or both positions of the triple.
• Situation: adding entities to a KG
• Transfer - predict links on entities that are unseen at training time.
• Situation: predicting links on an unseen KG

19.
Dynamic Transfer

20.
Representations for other tasks
https://github.com/dfdazac/blp

21.
“john lennon, parents”
“bicycle holiday nature”
“Airports in Germany”
“What is the longest river?”.

22.
Future: Sub-graph Prediction

23.
Future: Learning KG Pipelines End-to-End
Paul T. Groth, Antony Scerri, Ron Daniel, Bradley P. Allen:
End-to-End Learning for Answering Structured Queries Directly over Text. DL4KG@ESWC 2019: 57-70

24.
KG POPULATION &
THE LONG TAIL

25.
The Importance of Attributes

26.
• Extract numeric patterns (12 ± 3, 53–55, 0.245)
• Extract corresponding units (°C, μM, hours, h, MPa)
• Nanoamperes (nA) for neural cell Rheobase values
• Megapascals (MPa) for compressive strength of concrete
• Milligrams per Kilogram (mg/kg) for administered drug dosages
Units of Measurement Corey Harper

27.
Simple Annotation - Surprisingly Hard

28.
https://competitions.codalab.org/competitions/25770

29.
Task List

30.
Guidelines for applying the model

31.
More and more detailed guidelines...

32.
And started annotating, a paragraph at a time...

33.
NLP Competitions -- Labs Online Lecture -- Thorne, Harper
But the paragraphs got messy...
2020-12-07

34.
2020-12-07
https://github.com/harperco/measeval

35.
NLP Competitions -- Labs Online Lecture -- Thorne, Harper
2020-12-07
40

36.
NLP Competitions -- Labs Online Lecture -- Thorne, Harper
Some Inter-annotator Agreement Info (Krippendorff’s Alpha)
2020-12-07
41

37.
NLP Competitions -- Labs Online Lecture -- Thorne, Harper
Evaluation Output
2020-12-07
42

38.
NLP Competitions -- Labs Online Lecture -- Thorne, Harper
Evaluation Process
2020-12-07
43

39.
NLP Competitions -- Labs Online Lecture -- Thorne, Harper
Local Evaluation Micro Averages
• Can run micro averages
• Broken down:
• By entity / relation class
• By document (paragraph)
• By class & document
• Could add additional analysis
2020-12-07
44

40.
KG POPULATION &
THE LONG TAIL

41.
LINKING & IDENTITY
Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 2129–2137
Marseille, 11–16 May 2020
c European Language Resources Association (ELRA), licensed under CC-BY-NC
2129
Towards Entity Spaces
Marieke van Erp⇤
, Paul Groth†
⇤
KNAW Humanities Cluster - DHLab, Amsterdam, NL
†
University of Amsterdam, Amsterdam, NL
marieke.van.erp@dh.huc.knaw.nl, p.groth@uva.nl
Abstract
Entities are a central element of knowledge bases and are important input to many knowledge-centric tasks including text analysis.
For example, they allow us to find documents relevant to a specific entity irrespective of the underlying syntactic expression within a
document. However, the entities that are commonly represented in knowledge bases are often a simplification of what is truly being
referred to in text. For example, in a knowledge base, we may have an entity for Germany as a country but not for the more fuzzy concept
of Germany that covers notions of German Population, German Drivers, and the German Government. Inspired by recent advances in
contextual word embeddings, we introduce the concept of entity spaces - specific representations of a set of associated entities with
near-identity. Thus, these entity spaces provide a handle to an amorphous grouping of entities. We developed a proof-of-concept for
English showing how, through the introduction of entity spaces in the form of disambiguation pages, the recall of entity linking can be
improved.
Keywords: entity, identity, knowledge representation, entity linking
1. Introduction
Entities are a central element for knowledge bases and
text analysis tasks (Balog, 2018). However, the way in
which entities are represented in knowledge bases and
how subsequent tools use these representations are a sim-
plification of the complexity of many entities. For ex-
ample, the entity Germany in Wikidata as represented
by wikidata:Q183 focuses on its properties as a location
and geopolitical entity due to its membership as an in-
stance of sovereign state, country, federal state, repub-
lic, social state, legal state, and administrative territorial
entity. Similarly, in DBpedia (version 2016-10), Ger-
many is represented as entity of type populated place and
some subtypes such as yago:WikicatFederalCountries and
yago:WikicatMemberStatesOfTheEuropeanUnion.1
However, when the term Germany is used in text, it can
take on many meanings that all have ‘something to do’ with
Germany as it is represented in knowledge bases, but are all
not quite the same:
(1) Germany imported 47,600 sheep from Britain last
year, nearly half of total imports.
(2) German July car registrations up 14.2 pct yr / yr.
(3) Australia last won the Davis Cup in 1986, but they
were beaten finalists against Germany three years
ago under Fraser’s guidance.
In Example (1), Germany refers partly to the location, but
a location usually cannot take on an active role, such that
the entity ‘importing’ the sheep is most likely a referent to
the German meat industry. Germany in Example (2), refers
to the German population buying and registering more cars
than a year before. Finally, in Example (3), Germany refers
to the German Davis Cup team from 1993 (the news article
is from 1996). In the AIDA-YAGO dataset, this entity is
tagged as dbp:Germany Davis Cup team but this presents
1
Germany also has rdf:type dbo:Person but we assume this
is a glitch.
us with another layer of identities, namely that every year,
or every couple of years, the German Davis cup team con-
sists of different players. In 1993, the German Davis cup
team consisted of Michael Stich and Marc-Kevin Goellner,
in 1996 of David Prinosil and Hendrik Dreekmann and at
the time of writing this article in 2019 of Alexander Zverev
and Philipp Kohlschreiber. Both MAG (Moussallem et al.,
2017) and DBpedia spotlight (Daiber et al., 2013a) annotate
Australia and Germany in Example (3) as dbp:Australia
and dbp:Germany respectively. While both the annotations
and automatic linkages are close to the identity of the en-
tity in resolving these referents to dbp:Germany, we argue
this is an underspecification and highlights a larger problem
with identity representation in knowledge bases.
Collapsing of identities has been a frequent topic within
Semantic Web discourse. However, most discussions have
focused on issues with owl:sameAs links (McCusker and
McGuinness, 2010; Raad et al., 2018). However, the prob-
lem of simplified entity representations (e.g. the collaps-
ing of identities) also occurs before the creation of such
owl:sameAs links. Specifically, with the fact that most
knowledge bases represent a single or limited number of
an entity’s facets. In this paper, we analyse the extent of
the problem by connecting Semantic Web representations
of identity to linguistic representations of entities, namely
coreference and near-identity. To overcome this identity
problem, we argue for the introduction of explicit represen-
tations of near-identity within knowledge bases. We term
these explicit representations - entity spaces. We illustrate
how the introduction of entity spaces can boost the perfor-
mance of state-of-the-art entity linking pipelines.
Our contributions are: 1) the definition of entity spaces; 2)
a prototype showing the use of entity spaces over multiple
entity linking pipelines; and 3) experiments on 13 English
entity linking datasets showing the impact of a more toler-
ant approach to entity linking made possible through entity
spaces.
Our code and experimental results are available via https:
//github.com/MvanErp/entity-spaces.
A good question:
Thinking about an answer:

42.
Situated Dialogue
• Computer mediated
dialogue is omnipresent
• Key observation: dialogue is
situated. It not only is about
the “back-and-forth”
between parties but also the
larger environment (e.g.
documents, concepts,
projects, world knowledge).
• Not just chat

43.
Complex Environments (e.g. Standards
Organizations)
in-sight-it.github.io

44.
conversationkg - kg extraction from dialogue

45.
Concept1
Concept2 Concept3
KOS
Professional
Curators
Literature
Software
Non-professional
contributors
1. dealing with changing cultural and societal
norms, specifically to address or correct bias;
2. political influence
3. new concepts and terminology arising from
discoveries or change in perspective within a
technical/scientific community
4. gardening
5. incremental contributorship
6. progressive formalization
7. software and automation
8. integration of large numbers of data sources
9. variance in algorithm training data
Data
⚐
Society & Politics
(4, 5, 6)
(7, 8, 9)
(3)
(1, 2)
Source:
Michael Lauruhn and Paul Groth.
“Sources of Change for Modern Knowledge Organization Systems." Knowledge Organization 43, no. 8 (2016).

46.
Knowledge Engineering Revisited
• Knowledge graphs are built ad-hoc
• 100s of components (extractors, scrapers, quality,
scoring, user feedback, ….)
• Unique for each organization
• Existing knowledge engineering theory does not apply:
• Assumes small scale
• Assumes slow change
• People-centric
• Expressive representations
• an updated theory and methods for knowledge
engineering designed for the demands of modern
knowledge graphs

47.
knowledgescientist.org

48.
Conclusion
• Knowledge graphs require maintenance
• Maintenance is frequently people work
• Need for new ML based methods & new human + machine work
fl
ows
• Interested? Happy to talk more
Paul Groth | @pgroth | pgroth.com | indelab.org