HAL Id: hal-01472497
https://hal.archives-ouvertes.fr/hal-01472497
Submitted on 20 Feb 2017
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of sci-
entic research documents, whether they are pub-
lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diusion de documents
scientiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
Yago: A Core of Semantic Knowledge Unifying
WordNet and Wikipedia
Fabian Suchanek, Gjergji M Kasneci, Gerhard M Weikum
To cite this version:
Fabian Suchanek, Gjergji M Kasneci, Gerhard M Weikum. Yago: A Core of Semantic Knowledge
Unifying WordNet and Wikipedia. 16th international conference on World Wide Web, May 2007,
Ban, Canada. pp.697 - 697, �10.1145/1242572.1242667�. �hal-01472497�
YAGO: A Core of Semantic Knowledge
Unifying WordNet and Wikipedia
Fabian M. Suchanek
Max-Planck-Institut
Saarbr
¨
ucken / Germany
suchanekaOmpii.mpg.de
Gjergji Kasneci
Max-Planck-Institut
Saarbr
¨
ucken / Germany
kasneciaOmpii.mpg.de
Gerhard Weikum
Max-Planck-Institut
Saarbr
¨
ucken / Germany
weikumaOmpii.mpg.de
ABSTRACT
We present YAGO, a light-weight and extensible ontology
with high coverage and quality. YAGO builds on entities
and relations and currently contains more than 1 million
entities and 5 million facts. This includes the Is-A hierarchy
as well as non-taxonomic relations between entities (such
as hasWonPrize). The facts have been automatically ex-
tracted from Wikipedia and unified with WordNet, using
a carefully designed combination of rule-based and heuris-
tic methods described in this paper. The resulting knowl-
edge base is a major step beyond WordNet: in quality by
adding knowledge ab out individuals like persons, organiza-
tions, products, etc. with their semantic relationships – and
in quantity by increasing the number of facts by more than
an order of magnitude. Our empirical evaluation of fact cor-
rectness shows an accuracy of about 95%. YAGO is based on
a logically clean model, which is decidable, extensible, and
compatible with RDFS. Finally, we show how YAGO can be
further extended by state-of-the-art information extraction
techniques.
Categories and Subject Descriptors
H.0 [Information Systems]: General
General Terms
Knowledge Extraction, Ontologies
Keywords
Wikip edia, WordNet
1. INTRODUCTION
1.1 Motivation
Many applications in modern information technology uti-
lize ontological background knowledge. This applies above
all to applications in the vision of the Semantic Web, but
there are many other application fields. Machine translation
(e.g. [5]) and word sense disambiguation (e.g. [3]) exploit
lexical knowledge, query expansion uses taxonomies (e.g.
[16, 11, 27]), document classification based on supervised or
semi-sup erv ised learning can be combined with ontologies
(e.g. [14]), and [13] demonstrates the utility of background
knowledge for question answering and information retrieval.
Copyright is held by the International World Wide Web Conference Com-
mittee (IW3C2). Distribution of these papers is limited to classroom use,
and personal use by others.
WWW 2007, May 8–12, 2007, Banff, Alberta, Canada.
ACM 978-1-59593-654-7/07/0005.
Furthermore, ontological knowledge structures play an im-
portant role in data cleaning (e.g., for a data warehouse) [6],
record linkage (aka. entity resolution) [7], and information
integration in general [19].
But the existing applications typically use only a single
source of background knowledge (mostly WordNet [10] or
Wikip edia). They could boost their performance, if a huge
ontology with knowledge from several sources was available.
Such an ontology would have to be of high quality, with ac-
curacy close to 100 percent, i.e. comparable in quality to
an encyclopedia. It would have to comprise not only con-
cepts in the style of WordNet, but also named entities like
people, organizations, geographic locations, books, songs,
pro ducts, etc., and also relations among these such as what-
is-located-where, who-was-born-when, who-has-won-which-
prize, etc. It would have to be extensible, easily re-usable,
and application-independent. If such an ontology were avail-
able, it could boost the performance of existing applications
and also open up the path towards new applications in the
Semantic Web era.
1.2 Related Work
Knowledge representation is an old field in AI and has
provided numerous models from frames and KL-ONE to
recent variants of description logics and RDFS and OWL
(see [22] and [24]). Numerous approaches have been pro-
posed to create general-purpose ontologies on top of these
representations. One class of approaches focuses on extract-
ing knowledge structures automatically from text corpora.
These approaches use information extraction technologies
that include pattern matching, natural-language parsing,
and statistical learning [25, 9, 4, 1, 23, 20, 8]. These tech-
niques have also been used to extend WordNet by Wikip edia
individuals [21]. Another project along these lines is Know-
ItAll [9], which aims at extracting and compiling instances
of unary and binary predicate instances on a very large scale
– e.g., as many soccer players as possible or almost all com-
pany/CEO pairs from the business world. Although these
approaches have recently improved the quality of their re-
sults considerably, the quality is still significantly below that
of a man-made knowledge base. Typical results contain
many false positives (e.g., IsA(Aachen Cathedral, City), to
give one example from KnowItAll). Furthermore, obtaining
a recall above 90 percent for a closed domain typically en-
tails a drastic loss of precision in return. Thus, information-
extraction approaches are only of little use for applications
that need near-perfect ontologies (e.g. for automated rea-
soning). Furthermore, they typically do not have an explicit
(logic-based) knowledge representation model.
Due to the quality bottleneck, the most successful and
widely employed ontologies are still man-made. These in-
clude WordNet [10], Cyc or OpenCyc [17], SUMO [18], and
especially domain-specific ontologies and taxonomies such as
SNOMED
1
or the GeneOntology
2
. These knowledge sources
have the advantage of satisfying the highest quality expecta-
tions, because they are manually assembled. However, they
suffer from low coverage, high cost for assembly and quality
assurance, and fast aging. No human-made ontology knows
the most recent Windows version or the latest soccer stars.
1.3 Contributions and Outline
This paper presents YAGO
3
, a new ontology that com-
bines high coverage with high quality. Its core is assem-
bled from one of the most comprehensive lexicons available
to day, Wikipedia. But rather than using information ex-
traction methods to leverage the knowledge of Wikipedia,
our approach utilizes the fact that Wikip edia has category
pages. Category pages are lists of articles that belong to a
specific category (e.g., Zidane is in the category of French
football players
4
). These lists give us candidates for enti-
ties (e.g. Zidane), candidates for concepts (e.g. IsA(Zidane,
FootballPlayer)) [15] and candidates for relations (e.g. isC-
itizenOf(Zidane, France)). In an ontology, concepts have to
be arranged in a taxonomy to be of use. The Wikipedia
categories are indeed arranged in a hierarchy, but this hier-
archy is barely useful for ontological purposes. For example,
Zidane is in the super-category named ”Football in France”,
but Zidane is a football player and not a football. WordNet,
in contrast, provides a clean and carefully assembled hierar-
chy of thousands of concepts. But the Wikipedia concepts
have no obvious counterparts in WordNet.
In this paper we present new techniques that link the
two sources with near-perfect accuracy. To the best of
our knowledge, our method is the first approach that ac-
complishes this unification between WordNet and facts de-
rived from Wikipedia with an accuracy of 97%. This al-
lows the YAGO ontology to profit, on one hand, from the
vast amount of individuals known to Wikipedia, while ex-
ploiting, on the other hand, the clean taxonomy of concepts
from WordNet. Currently, YAGO contains roughly 1 million
entities and 5 million facts about them.
YAGO is based on a data model of entities and binary re-
lations. But by means of reification (i.e., introducing iden-
tifiers for relation instances) we can also express relations
between relation instances (e.g., popularity rankings of pairs
of soccer players and their teams) and general properties of
relations (e.g., transitivity or acyclicity). We show that, de-
spite its expressiveness, the YAGO data model is decidable.
YAGO is designed to be extendable by other sources – be
it by other high quality sources (such as gazetteers of geo-
graphic places and their relations), by domain-specific ex-
tensions, or by data gathered through information extrac-
tion from Web pages. We conduct an enrichment experi-
ment with the state-of-the-art information extraction system
Leila[25]. We observe that the more facts YAGO contains,
the better it can be extended. We hypothesize that this pos-
itive feedback loop could even accelerate future extensions.
The rest of this paper is organized as follows. In Section
2 we introduce YAGO’s data model. Section 3 describes the
1
http://www.snomed.org
2
http://www.geneontology.org/
3
Yet Another Great Ontology
4
So ccer is called football in some countries
sources from which the current YAGO is assembled, namely,
Wikip edia and WordNet. In Section 4 we give an overview of
the system behind YAGO. We explain our extraction tech-
niques and we show how YAGO can be extended by new
data. Section 5 presents an evaluation, a comparison to
other ontologies, an enrichment experiment and sample facts
from YAGO. We conclude with a summary in Section 6.
2. THE YAGO MODEL
2.1 Structure
To accommodate the ontological data we already ex-
tracted and to be prepared for future extensions, YAGO
must be based on a thorough and expressive data model.
The model must be able to express entities, facts, relations
between facts and properties of relations. The state-of-the-
art formalism in knowledge representation is currently the
Web Ontology Language OWL [24]. Its most expressive vari-
ant, OWL-full, can express properties of relations, but is
undecidable. The weaker variants of OWL, OWL-lite and
OWL-DL, cannot express relations between facts. RDFS,
the basis of OWL, can express relations between facts, but
provides only very primitive semantics (e.g. it does not know
transitivity). This is why we introduce a slight extension of
RDFS, the YAGO model. The YAGO model can express
relations between facts and relations, while it is at the same
time simple and decidable.
As in OWL and RDFS, all objects (e.g. cities, people,
even URLs) are represented as entities in the YAGO model.
Two entities can stand in a relation. For example, to state
that Albert Einstein won the Nobel Prize, we say that the
entity Albert Einstein stands in the ha sWonPrize rela-
tion with the entity Nobel Prize. We write
AlbertEinstein hasWonPrize NobelPrize
Numbers, dates, strings and other literals are represented as
entities as well. This means that they can stand in relations
to other entities. For example, to state that Albert Einstein
was born in 1879, we write:
AlbertEinstein bornInYear 1879
Entities are abstract ontological objects, which are
language-independe nt in the ideal case. Language uses
words to refer to these entities. In the YAGO model, words
are entities as well. This makes it possible to express that a
certain word refers to a certain entity, like in the following
example:
”Einstein” means AlbertEinstein
This allows us to deal with synonymy and ambiguity. The
following line says that ”Einstein” may also refer to the mu-
sicologist Alfred Einstein:
”Einstein” means AlfredEinstein
We use quotes to distinguish words from other entities. Sim-
ilar entities are grouped into classes. For example, the class
physicist comprises all physicists and the class word com-
prises all words. Each entity is an instance of at least one
class. We express this by the type relation:
AlbertEinstein type physicist
Classes are also entities. Thus, each class is itself an instance
of a class, namely of the class class. Classes are arranged
in a taxonomic hierarchy, expressed by the subClassOf re-
lation:
physicist subClassOf scientist
In the YAGO model, relations are entities as well. This
makes it possible to represent properties of relations (like
transitivity or subsumption) within the model. The follow-
ing line, e.g., states that the subClassOf relation is tran-
sitive by making it an instance of the class transitive-
Relation:
subclassOf type transitiveRelation
The triple of an entity, a relation and an entity is called
a fact. The two entities are called the arguments of the
fact. Each fact is given a fact identifier. As RDFS, the
YAGO model considers fact identifiers to be entities as well.
This allows us to represent for example that a certain fact
was found at a certain URL. For example, suppose that the
ab ove fact (Albert Einstein, bornInYear, 1879) had the
fact identifier #1, then the following line would say that this
fact was found in Wikipedia:
#1 foundIn http : //www.wikipedia.org/Einstein
We will refer to entities that are neither facts nor relations
as common entities. Common entities that are not classes
will be called individuals. Then, a YAGO ontology over a
finite set of common entities C, a finite set of relation names
R and a finite set of fact identifiers I is a function
y : I → (I ∪ C ∪ R) × R × (I ∪ C ∪ R)
A YAGO ontology y has to be injective and total to ensure
that every fact identifier of I is mapped to exactly one fact.
Some facts require more than two arguments (for example
the fact that Einstein won the Nobel Prize in 1921). One
common way to deal with this problem is to use n-ary re-
lations (as for example in won-prize-in-year(Einstein,
Nobel-Prize, 1921)). In a relational database setting,
where relations correspond to tables, this has the disadvan-
tage that much space will be wasted if not all arguments of
the n-ary facts are known. Worse, if an argument (like e.g.
the place of an event) has not been foreseen in the design
phase of the database, the argument cannot be represented.
Another way of dealing with an n-ary relation is to intro-
duce a binary relation for each argument (e.g. winner,prize,
time). Then, an n-ary fact can be represented by a new
entity that is linked by these binary relations to all of its
arguments (as is proposed for OWL):
AlbertEinstein winner EinsteinWonNP1921
NobelPrize prize EinsteinWonNP1921
1921 time EinsteinWonNP1921
However, this method cannot deal with additional argu-
ments to relations that were designed to be binary. The
YAGO model offers a simple solution to this problem: It is
based on the assumption that for each n-ary relation, a pri-
mary pair of its arguments can be identified. For example,
for the above won-prize-in-year-relation, the pair of the
person and the prize could be considered a primary pair.
The primary pair can be represented as a binary fact with
a fact identifier:
#1 : AlbertEinstein hasWonPrize NobelPrize
All other arguments can be represented as relations that
hold between the primary pair and the other argument:
#2 : #1 time 1921
2.2 Semantics
This section will give a model-theoretic semantics to
YAGO. We first prescribe that the set of relation names R
for any YAGO ontology must contain at least the relation
names type, subClassOf, domain, range and subRelation-
Of. The set of common entities C must contain at least
the classes entity, class, relation, acyclicTransitive-
Relation and classes for all literals (as evident from the
following list). For the rest of the paper, we assume a given
set of common entities C and a given set of relations R.
The set of fact identifiers used by a YAGO ontology y is
implicitly given by I = domain(y). To define the semantics
of a YAGO ontology, we consider only the set of possible
facts F = (I ∪ C ∪ R) × R × (I ∪ C ∪ R). We define a
rewrite system → ⊆ P(F) × P(F), i.e. → reduces one
set of facts to another set of facts. We use the shorthand
notation {f
1
, ..., f
n
} → f to say that
F ∪ {f
1
, ..., f
n
} → F ∪ {f
1
, ..., f
n
} ∪ {f}
for all F ⊆ F, i.e. if a set of facts contains the facts f
1
, ..., f
n
,
then the rewrite rule adds f to this set. Our rewrite system
contains the following (axiomatic) rules:
5
∅ → (domain, domain, relation)
∅ → (domain, range, class)
∅ → (range, domain, relation)
∅ → (range, range, class)
∅ → (subClassOf, type, acyclicTransitiveRelation)
∅ → (subClassOf, domain, class)
∅ → (subClassOf, range, class)
∅ → (type, range, class)
∅ → (subRelationOf, type, acyclicTransitiveRelation)
∅ → (subRelationOf, domain, relation)
∅ → (subRelationOf, range, relation)
∅ → (boolean, subClassOf, literal)
∅ → (number, subClassOf, literal)
∅ → (rationalNumber, subClassOf, number)
∅ → (integer, subClassOf, rationalNumber)
∅ → (timeInterval, subClassOf, literal)
∅ → (dateTime, subClassOf, timeInterval)
∅ → (date, subClassOf, timeInterval)
∅ → (string, subClassOf, literal)
∅ → (character, subClassOf, string)
∅ → (word, subClassOf, string)
∅ → (URL, subClassOf, string)
Furthermore, it contains the following rules for all
r, r
1
, r
2
∈ R, x, y, c, c
1
, c
2
∈ I ∪ C ∪ R, r
1
6=
type, r
2
6= subRelationOf, r 6= subRelationOf,
r 6= type, c 6= acyclicTransitiveRelation, c
2
6=
acyclicTransitiveRelation:
(1) {(r
1
, subRelationOf, r
2
), (x, r
1
, y)} → (x, r
2
, y)
(2) {(r, type, acyclicTransitiveRelation), (x, r, y), (y, r, z)}
→ (x, r, z)
(3) {(r, domain, c), (x, r, c)} → (x, type, c)
(4) {(r, range, c), (x, r, y)} → (y, type, c)
(5) {(x, type, c
1
), (c
1
, subClassOf, c
2
)} → (x, type, c
2
)
Theorem 1: [Convergence of →]
Given a set of facts F ⊂ F, the largest set S with F →
∗
S
is unique.
(The theorems are proven in the appendix.) Given a YAGO
ontology y, the rules of → can be applied to its set of facts,
5
The class hierarchy of literals is inspired by SUMO[18]
r ange(y). We call the largest set that can be produced by
applying the rules of → the set of derivable facts of y, D(y).
Two YAGO ontologies y
1
, y
2
are equivalent if the fact iden-
tifiers in y
2
can be renamed so that
(y
1
⊆ y
2
∨ y
2
⊆ y
1
) ∧ D(y
1
) = D(y
2
)
The deductive closure of a YAGO ontology y is computed by
adding the derivable facts to y. Each derivable fact (x, r, y)
needs a new fact identifier, which is just f
x,r,y
. Using a
relational notation for the function y, we can write this as
y
∗
:= y ∪ { (f
r,x,y
, (r, x, y)) |
(x, r, y) ∈ D(y) , (r, x, y) 6∈ range(y) }
A structure for a YAGO ontology y is a triple of
• a set U (the universe)
• a function D : I ∪ C ∪ R → U (the denotation)
• a function E : D(R) → U ×U (the extension function)
Like in RDFS, a YAGO structure needs to define the exten-
sions of the relations by the extension function E. E maps
the denotation of a relation symbol to a relation on universe
elements. We define the interpretation Ψ with respect to a
structure < U, D, E > as the following relation:
Ψ := {(e
1
, r, e
2
) | (D(e
1
), D (e
2
)) ∈ E(D(r))}
We say that a fact (e
1
, r, e
2
) is true in a structure, if it
is contained in the interpretation. A model of a YAGO
ontology y is a structure such that
1. all facts of y
∗
are true
2. if Ψ(x, type, literal) for some x, then D(x) = x
3. if Ψ(r, type, acyclicTransitiveRelation) for some r,
then there exists no x such that Ψ(x, r, x)
A YAGO ontology y is called consistent iff there exists a
mo del for it. Obviously, a YAGO ontology is consistent iff
6 ∃x, r : (r, type, acyclicTransitiveRelation) ∈ D(y)
∧ (x, r, x) ∈ D(y)
Since D(y) is finite, the consistency of a YAGO ontology is
decidable. A base of a YAGO ontology y is any equivalent
YAGO ontology b with b ⊆ y. A canonical base of y is a
base so that there exists no other base with less elements.
Theorem 2: [Uniqueness of the Canonical Base]
The canonical base of a consistent YAGO ontology is unique.
In fact, the canonical base of a YAGO ontology can be com-
puted by greedily removing derivable facts from the ontol-
ogy. This makes the canonical base a natural choice to effi-
ciently store a YAGO ontology.
2.3 Relation to Other Formalisms
The YAGO model is very similar to RDFS. In RDFS, rela-
tions are called properties. Just as YAGO, RDFS knows the
properties domain, range, subClassOf and subPropertyOf
(i.e. subRelationOf). These properties have a semantics
that is equivalent to that of the corresponding YAGO re-
lations. RDFS also knows fact identifiers, which can occur
as arguments of other facts. The following excerpt shows
how some sample facts of Section 2.1 can be represented in
RDFS. Each fact of YAGO becomes a triple in RDFS.
<rdf:Description
rdf:about="http://mpii.mpg.de/yago#Albert_Einstein">
<yago:bornInYear rdf:ID="f1">1879</yago:bornInYear>
</rdf:Description>
<rdf:Description
rdf:about="http://mpii.mpg.de/yago#f1">
<yago:foundIn rdf:ID="f2" rdf:resource="http:..."/>
</rdf:Description>
However, RDFS does not have a built-in transitive relation
or an acyclic transitive relation, as YAGO does. This en-
tails that the property acyclicTransitiveRelation can be
defined and used, but that RDFS would not know its se-
mantics.
YAGO uses fact identifiers, but it does not have built-
in relations to make logical assertions about facts (e.g. it
do e s not allow to say that a fact is false). If one relies on
the denotation to map a fact identifier to the corresponding
fact element in the universe, one can consider fact identi-
fiers as simple individuals. This abandons the syntactic link
between a fact identifier and the fact. In return, it opens
up the possibility of mapping a YAGO ontology to an OWL
ontology under certain conditions. OWL has built-in coun-
terparts for almost all built-in data types, classes, and rela-
tions of YAGO. The only concept that does not have an ex-
act built-in counterpart is the acyclicTransitiveRelation.
However, this is about to change. OWL is currently being
refined to its successor, OWL 1.1. The extended description
logic SROIQ [12], which has been adopted as the logical
basis of OWL 1.1, allows to express irreflexivity and transi-
tivity. This allows to define acyclic transitivity. We plan to
investigate the relation of YAGO and OWL once OWL 1.1
has been fully established.
3. SOURCES FOR YAGO
3.1 WordNet
WordNet is a semantic lexicon for the English language
developed at the Cognitive Science Laboratory of Prince-
ton University. WordNet distinguishes between words as
literally appearing in texts and the actual senses of the
words. A set of words that share one sense is called a
synset. Thus, each synset identifies one sense (i.e., se-
mantic concept). Words with multiple meanings (ambigu-
ous words) belong to multiple synsets. As of the current
version 2.1, WordNet contains 81,426 synsets for 117,097
unique nouns. (Wordnet also includes other types of words
like verbs and adjectives, but we consider only nouns in
this paper.) WordNet provides relations between synsets
such as hypernymy/hyponymy (i.e., the relation between a
sub-concept and a super-concept) and holonymy/meronymy
(i.e., the relation b etween a part and the whole); for this
pap e r, we focus on hypernyms/hyponyms. Conceptually,
the hyp ernymy relation in WordNet spans a directed acyclic
graph (DAG) with a single source node called Entity.
3.2 Wikipedia
Wikip edia is a multilingual, Web-based encyclopedia. It is
written collaboratively by volunteers and is available for free.
We downloaded the English version of Wikipedia in January
2007, which comprised 1,600,000 articles at that time. Each
Wikip edia article is a single Web page and usually describes
a single topic.
The majority of Wikipedia pages have been manually as-
signed to one or multiple categories. The page about Albert