See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/323316736

Towards a Definition of Knowledge Graphs

Conference Paper · September 2016

CITATIONS READS
524 16,268

2 authors:

Lisa Ehrlinger Wolfram Wöß

Software Competence Center Hagenberg Johannes Kepler University Linz
39 PUBLICATIONS 863 CITATIONS 99 PUBLICATIONS 1,372 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Wolfram Wöß on 21 February 2018.

The user has requested enhancement of the downloaded file.

 Towards a Definition of Knowledge Graphs

Lisa Ehrlinger and Wolfram Wöß

Institute for Application Oriented Knowledge Processing
Johannes Kepler University Linz, Austria
{lisa.ehrlinger | wolfram.woess}@jku.at

ABSTRACT source for several prominent knowledge representation ap-

Recently, the term knowledge graph has been used frequently plications. Other definitions may lead to the assumption
in research and business, usually in close association with that knowledge graph is a synonym for any graph-based
Semantic Web technologies, linked data, large-scale data knowledge representation (cf. [12, 16]). We argue that such
analytics and cloud computing. Its popularity is clearly in- a definition is not enough for an adequate application of
fluenced by the introduction of Google’s Knowledge Graph knowledge graphs, since it does not enforce a minimum set
in 2012, and since then the term has been widely used with- of requirements a KG has to fulfill. Thus, even a simple
out a definition. A large variety of interpretations has ham- graph-based vocabulary could be published as knowledge
pered the evolution of a common understanding of knowledge graph. In addition, such a definition creates an entrance
graphs. Numerous research papers refer to Google’s Knowl- barrier for people who are unfamiliar with knowledge graphs
edge Graph, although no official documentation about the and want to delve deeper into the topic or aim at building
used methods exists. The prerequisite for widespread aca- a KG on their own. A clear definition propagates a shared
demic and commercial adoption of a concept or technology is understanding of benefits, improvements, or drawbacks that
a common understanding, based ideally on a definition that can be expected if someone builds a knowledge graph. Thus,
is free from ambiguity. We tackle this issue by discussing we provide a discussion of knowledge graphs and motivate a
and defining the term knowledge graph, considering its his- definition to support a common understanding.
tory and diversity in interpretations and use. Our goal is to This paper is organized as follows: Section 2 presents re-
propose a definition of knowledge graphs that serves as basis lated state-of-the-art research and existing attempts at defin-
for discussions on this topic and contributes to a common ing knowledge graphs. Section 3 provides a short overview of
vision. historic and current knowledge graph applications and Sec-
tion 4 delimits and differentiates the term knowledge graph
from similar terms and introduces our definition.
CCS Concepts
•Information systems → Data management systems; In- 2. SELECTED DEFINITIONS
formation systems applications; Knowledge graphs have been in the focus of research since
2012 resulting in a wide variety of published descriptions
Keywords and definitions. Table 1 lists representative definitions and
Knowledge Graphs, Knowledge Bases, Ontologies, Knowledge demonstrates the lack of a common core, a fact that is also
Representation, Semantic Web. indicated by Paulheim [16] in 2015. Paulheim listed in his
survey of knowledge graph refinement the minimum set of
characteristics that must be present to distinguish knowledge
1. INTRODUCTION graphs from other knowledge collections (cf. first definition
Considerable research into knowledge graphs (KGs) has in Table 1), which basically restricts the term to any graph-
been carried out in recent years, especially in the Semantic based knowledge representation. In the online reviewing
Web community, and thus a variety of partially contradicting process of “Knowledge Graph Refinement: A Survey of Ap-
definitions and descriptions has emerged. The often quoted proaches and Evaluation Methods” [16], Noy1 agreed that a
blog entry by Google [18] basically describes an enhancement more precise definition was hard to find at that point. This
of their search engine with semantics. And also Wikipedia, statement points out the demand for closer investigation and
the most comprehensive encyclopedia in the web, does not deeper reflection in this area.
provide information about knowledge graphs in general, but Vague descriptions of knowledge graphs were published in
refers to the implementation by Google without mentioning the announcement of a special issue on knowledge graphs
the existence of other knowledge graphs. Although Wikipedia by the Journal of Web Semantics and by the Semantic Web
is no scientific reference source, it contributes to a common Company (cf. second and third definitions in Table 1). Both
understanding through its role as a primary information definitions could equally well describe an ontology or – even
more generally – any kind of semantic knowledge representa-
c 2016 Copyright held by the author/owner(s). 1
SEMANTICS 2016: Posters and Demos Track http://www.semantic-web-journal.net/content/
September 13-14, 2016, Leipzig, Germany knowledge-graph-refinement-survey-approaches-and-
evaluation-methods [August, 2016]
Definition Source
“A knowledge graph (i) mainly describes real world entities and their interrelations, organized Paulheim [16]
in a graph, (ii) defines possible classes and relations of entities in a schema, (iii) allows for
potentially interrelating arbitrary entities with each other and (iv) covers various topical
domains.”
“Knowledge graphs are large networks of entities, their semantic types, properties, and Journal of Web Semantics [12]
relationships between entities.”
“Knowledge graphs could be envisaged as a network of all kind things which are relevant Semantic Web Company [3]
to a specific domain or to an organization. They are not limited to abstract concepts and
relations but can also contain instances of things like documents and datasets.”
“We define a Knowledge Graph as an RDF graph. An RDF graph consists of a set of RDF Färber et al. [7]
triples where each RDF triple (s, p, o) is an ordered set of the following RDF terms: a
subject s ∈ U ∪ B, a predicate p ∈ U , and an object U ∪ B ∪ L. An RDF term is either a
URI u ∈ U , a blank node b ∈ B, or a literal l ∈ L.”
“[...] systems exist, [...], which use a variety of techniques to extract new knowledge, in the Pujara et al. [17]
form of facts, from the web. These facts are interrelated, and hence, recently this extracted
knowledge has been referred to as a knowledge graph.”

Table 1: Selected definitions of knowledge graph

tion and do not even enforce a graph structure. In addition, In the 1980s, researchers from the University of Groningen
size is highlighted as an essential characteristic, which is re- and the University of Twente in the Netherlands initially
flected by phrases such as “large networks” or “vast networks” introduced the term knowledge graph to formally describe
[11], while it remains unclear what “large” means in this con- their knowledge-based system that integrates knowledge from
text. Färber et al. defined a knowledge graph as an Resource different sources for representing natural language [10, 15].
Description Framework (RDF) graph and stated that the The authors proposed KGs with a limited set of relations and
term KG was coined by Google to describe any graph-based focus on qualitative modeling including human interaction,
knowledge base (KB) [7]. Although this definition is the only which clearly contrasts with the idea of KGs that has been
formal one, it contradicts with more general definitions as it widely discussed in recent years.
explicitly requires the RDF data model. Pujara et al. did In 2012, Google introduced the Knowledge Graph as a
not provide a concise definition, but rather described the semantic enhancement of Google’s search function that does
characteristics of knowledge graphs. Unlike the other defini- not match strings, but enables searching for “things”, in other
tions, which focus solely on the inner structure of the KG, words, real-world objects [18]. Although the blog post does
they highlighted the importance of an automatic extraction not provide any implementation details, it has been cited
system. In the preface of the 13th International Semantic more than 100 times according to Google Scholar3 . Since
Web Conference Proceedings (2014), the following statement 2012, the term knowledge graph is also used to describe a
was published: family of applications. Frequently mentioned implementa-
tions are DBPedia, YAGO (Yet Another Great Ontology),
Significantly, major companies, such as Google, Freebase, Wikidata, Yahoo’s semantic search assistant tool
Yahoo, Microsoft, and Facebook, have created their Spark, Google’s Knowledge Vault, Microsoft’s Satori and
own “knowledge graphs” that power semantic searches Facebook’s entity graph [7, 14, 16, 11]. Those applications
and enable smarter processing and delivery of differ in their characteristics, such as architecture, opera-
data: The use of these knowledge graphs is now tional purpose, and technology used, which makes it difficult
the norm rather than the exception. [14] to find a consensus and to create a definition of knowledge
graph. The lowest common denominator of the listed open
Once again, this highlights the demand for a common source applications is their use of Linked Data, whereas
definition, because it is necessary to define and differentiate hardly any proven information is available about Satori and
KGs from other concepts in order to make valuable and accu- the entity graph.
rate statements about the introduction and dissemination of In addition, the more specific term enterprise knowledge
knowledge graphs. Furthermore, this ISWC statement pro- graph is used by a few smaller companies, for example,
claims the use of knowledge graphs to be the norm in general, SindiceTech4 and the Semantic Web Company [3]. Both
instead of restricting the scope, domain, or application area companies seek to describe a similar model that extracts and
where KGs can be used beneficially and efficiently. Despite stores diverse enterprise data in a triple store and analyzes it
its lack of clarity, this statement seems to have inspired many by using machine learning techniques in order to acquire new
researchers to submit papers about knowledge graphs in the knowledge from the data and to reuse it in other applications.
following conference in 20152 .

3. KNOWLEDGE GRAPH APPLICATIONS 3

https://scholar.google.at/scholar?q=Introducing+the+
Knowledge+Graph%3A+things%2C+not+strings&btnG=
2 &hl=en&as sdt=0%2C5 [August, 2016]
http://iswc2015.semanticweb.org/program/accepted-
4
papers [August, 2016] http://www.sindicetech.com/overview.html [August, 2016]
4. TERMINOLOGICAL ANALYSIS AND The second interpretation leads to the assumption that a
DEFINITION knowledge graph is a knowledge-based system that contains a
knowledge base and a reasoning engine. Focusing on existing
When analyzing current research work that defines or ad- automatically generated “knowledge graphs”, we can identify
dresses knowledge graphs, two fundamental issues can be a further essential characteristic: collection, extraction, and
identified: (a) Google’s blog entry about their Knowledge integration of information from external sources extends a
Graph is cited as if it provides a proper explanation for con-
pure knowledge-based system with the concept of integration
stituting a knowledge graph (cf. [17, 19]), and (b) the terms systems. Most open source applications listed in Section 3
knowledge graph and knowledge base are used interchange- implement the integration aspect with Linked Data.
ably (cf. [5, 7, 8, 13, 16, 20]). The second problem leads to
the misleading assumption that the term knowledge graph is
Knowledge-based system
a synonym for knowledge base, which is itself often used as Source 1
synonym for ontology. An example of such confusion is that Knowledge base Reasoning
Source 2
both Knowledge Vault and Google’s Knowledge Graph have (e.g., ontology) engine
been called large-scale knowledge base by their respective cre- …
ators [5]. A further example is YAGO, which is – according
to its name – an ontology, but is referred to both as knowl- Figure 1: Architecture of a knowledge graph
edge base (cf. [5, 16]) and as knowledge graph (cf. [8, 21]).
Similarly, Yahoo employees [2] do not distinguish clearly be- Figure 1 illustrates the combination of these assumptions,
tween knowledge base, knowledge graph and ontology. They which yields an abstract knowledge graph architecture. Based
state that they construct their knowledge base by aligning on this architecture and derived from the terminological
new entities, relations and information with their common analysis, we define a knowledge graph as follows:
ontology. Therefore, incomplete, inconsistent and possibly
inaccurate information is turned into a rich, unified, disam- A knowledge graph acquires and integrates infor-
biguated knowledge graph. Based on this information, their mation into an ontology and applies a reasoner
understanding of a knowledge graph is the cleaned knowledge to derive new knowledge.
base that is the population (e.g., instances) of their ontology.
This definition aligns with the assumption that a knowl-
In order to distinguish between the terms, they must be
edge graph is somehow superior and more complex than
clarified explicitly. According to Akerkar and Sajja [1], a
a knowledge base (e.g., an ontology) because it applies a
knowledge-based system uses artificial intelligence to solve
reasoning engine to generate new knowledge and integrates
problems, and it consists of two parts: a knowledge base and
one or more information sources. Consequently, a manually
an inference engine. The knowledge base is a dataset with for-
created knowledge graph that does not support integration
mal semantics that can contain different kinds of knowledge,
aspects is a plain knowledge base or knowledge-based system
for example, rules, facts, axioms, definitions, statements, and
if it provides reasoning capabilities. The definition does not
primitives [4]. Thus, Knowledge Vault cannot be classified
take the quantity aspect (size) into account, especially with
as a true knowledge base, because it extends the idea of a
respect to a large ABox of the ontology, since it is not clear
pure semantic store with reasoning capabilities and therefore
what can be considered “large”. Instead, reasoning capabil-
bears more resemblance to a knowledge-based system.
ities are highlighted as an essential characteristic to derive
An ontology is as a formal, explicit specification of a shared
new knowledge and differentiate a KG from knowledge bases.
conceptualization that is characterized by high semantic ex-
Furthermore, the question arises what constitutes the dif-
pressiveness required for increased complexity [9]. Ontolog-
ference between the Semantic Web and knowledge graphs.
ical representations allow semantic modeling of knowledge,
Smaller KGs, for example, enterprise knowledge graphs, can
and are therefore commonly used as knowledge bases in artifi-
be clearly differentiated from the Semantic Web because of
cial intelligence (AI) applications, for example, in the context
their restricted domain. The goal of large search engines is
of knowledge-based systems. Application of an ontology as
to crawl and process all available information in the web,
knowledge base facilitates validation of semantic relationships
which leads to an increased interest in the widespread imple-
and derivation of conclusions from known facts for inference
mentation of semantic technology. Hardly any information is
(i.e., reasoning) [9]. We explicitly emphasize that an ontology
available on the technologies applied in Google’s Knowledge
does not differ from a knowledge base, although ontologies
Graph and Microsoft’s Satori, but Yahoo’s Spark and the
are sometimes erroneously classified as being at the same
Knowledge Vault apparently use Semantic Web standards
level as database schemas [6]. In fact, an ontology consists
such as RDF. Considering the layers of the Semantic Web,
not only of classes and properties (e.g., owl:ObjectProperty
a knowledge graph, in comparison, deploys either exactly
and owl:DatatypeProperty), but can also hold instances
the same technology for every layer or a similar one that
(i.e., the population of the ontology).
offers the same features. For example, companies might use
On the one hand, size is often mentioned as an essential
proprietary identifiers in place of URIs and JSON-LD5 as
characteristic of knowledge graphs, therefore a KG could be
serialization format substituting XML and RDF. However,
described as very large ontology. However, other contribu-
these technologies are just examples, and particularly in the
tors have pointed out that knowledge graphs are somehow
syntax layer XML is often replaced with more lightweight
superior to ontologies [3] and provide additional features.
and more easily readable formats such as Turtle, N-Triples
Thus, the difference between a knowledge graph and an on-
or N-Quads in the Semantic Web community. In conclusion,
tology could be interpreted either as a matter of quantity
the Semantic Web could be interpreted as the most com-
(e.g., a large ontology), or of extended requirements (e.g., a
prehensive knowledge graph, or – conversely – a knowledge
built-in reasoner that allows new knowledge to be derived).
5
http://json-ld.org [August, 2016]
 graph that crawls the entire web could be interpreted as and 1st Workshop on Data Science: Methods,
self-contained Semantic Web. Technology and Applications (DSci15), pages 18–21.
CEUR Workshop Proceedings, 2015.
5. CONCLUSION [9] C. Feilmayr and W. Wöß. An Analysis of Ontologies
Graph-based knowledge representation has been researched and their Success Factors for Application to Business.
for decades and the term knowledge graph does not consti- Data & Knowledge Engineering, 101:1–23, 2016.
tute a new technology. Rather, it is a buzzword reinvented [10] P. James. Knowledge Graphs. In Linguistic
by Google and adopted by other companies and academia Instruments in Knowledge Engineering, pages 97–117.
to describe different knowledge representation applications. Elsevier Science Publishers B.V., 1992.
We have proposed a definition of knowledge graph in order [11] S. Krause, L. Hennig, A. Moro, D. Weißenborn, F. Xu,
to promote a discussion and a common vision for further H. Uszkoreit, and R. Navigli. Sar-graphs: A Language
work in this area. There are essential differences in the Resource Connecting Linguistic Knowledge with
way knowledge representation applications (cf. Section 3) Semantic Relations from Knowledge Graphs. Journal
are built, ranging from completely handcrafted knowledge of Web Semantics: Science, Services and Agents on the
bases to automatically extracted and processed knowledge World Wide Web, Special Issue on Knowledge
graphs. Consequently, the term knowledge graph is not suit- Graphs(C):112–131, Mar. 2016.
able for describing all of these applications and should be [12] M. Kroetsch and G. Weikum. Journal of Web
used more carefully. Several applications need not be called Semantics: Special Issue on Knowledge Graphs.
knowledge graphs, because the terms knowledge base and http://www.websemanticsjournal.org/index.php/ps/
ontology describe them sufficiently and more accurately. Tak- announcement/view/19 [August, 2016].
ing into account the diverse applications, a KG bears more [13] Y. Ma, P. A. Crook, R. Sarikaya, and E. Fosler-Lussier.
resemblance to an abstract framework than to a mathemati- Knowledge Graph Inference for Spoken Dialog Systems.
cal structure. Our ongoing research focuses on an in-depth In Proceedings of 40th IEEE International Conference
analysis of our definition with respect to existing KG im- on Acoustics, Speech and Signal Processing (ICASSP)
plementations as well as the assessment of data quality in 2015. IEEE - Institute of Electrical and Electronics
knowledge graphs and their accessed sources. Engineers, April 2015.
[14] P. Mika, T. Tudorache, A. Bernstein, C. Welty, C. A.
6. REFERENCES Knoblock, D. Vrandečić, P. T. Groth, N. F. Noy,
[1] R. Akerkar and P. Sajja. Knowledge-Based Systems. K. Janowicz, and C. A. Goble, editors. The Semantic
Jones and Bartlett Publishers, USA, 1st edition, 2009. Web - ISWC 2014 - 13th International Semantic Web
[2] R. Blanco, B. B. Cambazoglu, P. Mika, and N. Torzec. Conference, Riva del Garda, Italy, October 19-23, 2014.
Entity Recommendations in Web Search. In Proceedings, Part I, volume 8796 of Lecture Notes in
Proceedings of the 12th International Semantic Web Computer Science. Springer, 2014.
Conference - Part II, ISWC ’13, pages 33–48, New [15] S. Nurdiati and C. Hoede. 25 Years Development of
York, USA, 2013. Springer. Knowledge Graph Theory: The Results and the
[3] A. Blumauer. From Taxonomies over Ontologies to Challenge, September 2008.
Knowledge Graphs, July 2014. https://blog.semantic- [16] H. Paulheim. Knowledge Graph Refinement: A Survey
web.at/2014/07/15/from-taxonomies-over-ontologies- of Approaches and Evaluation Methods. Semantic Web
to-knowledge-graphs [August, 2016]. Journal, (Preprint):1–20, 2016.
[4] J. Davies, R. Studer, and P. Warren. Semantic Web [17] J. Pujara, H. Miao, L. Getoor, and W. Cohen.
Technologies: Trends and Research in Ontology-based Knowledge Graph Identification. In Proceedings of the
Systems. John Wiley & Sons, 2006. 12th International Semantic Web Conference - Part I,
[5] X. Dong, E. Gabrilovich, G. Heitz, W. Horn, N. Lao, ISWC ’13, pages 542–557, New York, USA, 2013.
K. Murphy, T. Strohmann, S. Sun, and W. Zhang. Springer.
Knowledge Vault: A Web-scale Approach to [18] A. Singhal. Introducing the Knowledge Graph: Things,
Probabilistic Knowledge Fusion. In Proceedings of the not Strings, May 2012. https://googleblog.blogspot.co.
20th ACM SIGKDD International Conference on at/2012/05/introducing-knowledge-graph-things-not.
Knowledge Discovery and Data Mining, KDD ’14, html [August, 2016].
pages 601–610, New York, USA, 2014. ACM. [19] T. Steiner, R. Verborgh, R. Troncy, J. Gabarró Vallés,
[6] J. Euzenat and P. Shvaiko. Ontology Matching. and R. Van de Walle. Adding Realtime Coverage to the
Springer, Secaucus, NJ, USA, 2007. Google Knowledge Graph. In Poster and Demo
[7] M. Färber, B. Ell, C. Menne, A. Rettinger, and Proceedings of the 11th International Semantic Web
F. Bartscherer. Linked Data Quality of DBpedia, Conference, Nov. 2012.
Freebase, OpenCyc, Wikidata, and YAGO. Semantic [20] F. M. Suchanek and G. Weikum. Knowledge Bases in
Web Journal, 2016. http://www.semantic-web-journal. the Age of Big Data Analytics. Proceedings of the
net/content/linked-data-quality-dbpedia-freebase- VLDB Endowment, 7(13):1713–1714, Aug. 2014.
opencyc-wikidata-and-yago [August, 2016] (revised [21] A. Tonon, M. Catasta, R. Prokofyev, G. Demartini,
version, under review). K. Aberer, and P. Cudré-Mauroux. Contextualized
[8] M. Färber and A. Rettinger. A Statistical Comparison Ranking of Entity Types Based on Knowledge Graphs.
of Current Knowledge Bases. In Joint Proceedings of Journal of Web Semantics: Science, Services and
the Posters and Demos Track of 11th International Agents on the World Wide Web, Special Issue on
Conference on Semantic Systems - SEMANTiCS2015 Knowledge Graphs(C):170–183, Mar. 2016.

View publication stats