PreprintPDF Available

Information Extraction From Co-Occurring Similar Entities

Preprints and early-stage research may not have been peer reviewed yet.

Abstract and Figures

Knowledge about entities and their interrelations is a crucial factor of success for tasks like question answering or text summarization. Publicly available knowledge graphs like Wikidata or DBpedia are, however, far from being complete. In this paper, we explore how information extracted from similar entities that co-occur in structures like tables or lists can help to increase the coverage of such knowledge graphs. In contrast to existing approaches, we do not focus on relationships within a listing (e.g., between two entities in a table row) but on the relationship between a listing's subject entities and the context of the listing. To that end, we propose a descriptive rule mining approach that uses distant supervision to derive rules for these relationships based on a listing's context. Extracted from a suitable data corpus, the rules can be used to extend a knowledge graph with novel entities and assertions. In our experiments we demonstrate that the approach is able to extract up to 3M novel entities and 30M additional assertions from listings in Wikipedia. We find that the extracted information is of high quality and thus suitable to extend Wikipedia-based knowledge graphs like DBpedia, YAGO, and CaLiGraph. For the case of DBpedia, this would result in an increase of covered entities by roughly 50%.
Content may be subject to copyright.
Information Extraction From Co-Occurring Similar Entities
Nicolas Heist
Data and Web Science Group
University of Mannheim, Germany
Heiko Paulheim
Data and Web Science Group
University of Mannheim, Germany
Knowledge about entities and their interrelations is a crucial factor
of success for tasks like question answering or text summarization.
Publicly available knowledge graphs like Wikidata or DBpedia
are, however, far from being complete. In this paper, we explore
how information extracted from similar entities that co-occur in
structures like tables or lists can help to increase the coverage of
such knowledge graphs. In contrast to existing approaches, we do
not focus on relationships within a listing (e.g., between two entities
in a table row) but on the relationship between a listing’s subject
entities and the context of the listing. To that end, we propose a
descriptive rule mining approach that uses distant supervision to
derive rules for these relationships based on a listing’s context.
Extracted from a suitable data corpus, the rules can be used to
extend a knowledge graph with novel entities and assertions. In our
experiments we demonstrate that the approach is able to extract up
to 3M novel entities and 30M additional assertions from listings in
Wikipedia. We nd that the extracted information is of high quality
and thus suitable to extend Wikipedia-based knowledge graphs
like DBpedia, YAGO, and CaLiGraph. For the case of DBpedia, this
would result in an increase of covered entities by roughly 50%.
Information systems Information extraction
;Data extrac-
tion and integration;Association rules.
Entity co-occurrence, Information extraction, Novel entity detec-
tion, CaLiGraph, DBpedia
ACM Reference Format:
Nicolas Heist and Heiko Paulheim. 2021. Information Extraction From
Co-Occurring Similar Entities. In Proceedings of the Web Conference 2021
(WWW ’21), April 19–23, 2021, Ljubljana, Slovenia. ACM, New York, NY,
USA, 11 pages.
1.1 Motivation and Problem
In tasks like question answering, text summarization, or entity
disambiguation, it is essential to have background information
about the involved entities. With entity linking tools like DBpedia
Spotlight [
] or Falcon [
], one can easily identify named entities
This paper is published under the Creative Commons Attribution 4.0 International
(CC-BY 4.0) license. Authors reserve their rights to disseminate the work on their
personal and corporate Web sites with the appropriate attribution.
WWW ’21, April 19–23, 2021, Ljubljana, Slovenia
2021 IW3C2 (International World Wide Web Conference Committee), published
under Creative Commons CC-BY 4.0 License.
ACM ISBN 978-1-4503-8312-7/21/04.
Figure 1: Simplied view on the Wikipedia page of Gilby
Clarke with a focus on its title, sections, and listings.
in text and retrieve the respective entity in a background entity hub
of the linking tool (e.g. in a wiki like Wikipedia or in a knowledge
graph like DBpedia [
]). This is, however, only possible if the
entity in question is contained in the respective entity hub [29].
The trend of entities added to publicly available knowledge
graphs in recent years indicates that they are far from being com-
plete. The number of entities in Wikidata [
], for example, grew
by 37% in the time from October 2019 (61.7M) to October 2020
(84.5M). In the same time, the number of statements increased by
41% from 770M to 1085M.
According to [
], Wikidata describes
the largest number of entities and comprises – in terms of entities –
other open knowledge graphs to a large extent. Consequently, this
problem applies to all public knowledge graphs, and particularly
so for long-tail and emerging entities [6].
Automatic information extraction approaches can help mitigat-
ing this problem if the approaches can make sure that the extracted
information is of high quality. While the performance of open in-
formation extraction systems (i.e. systems that extract information
from general web text) has improved in recent years [
], the
quality of extracted information has not yet reached a level where
an integration into knowledge graphs like DBpedia should be done
without further ltering.
The extraction of information from semi-structured data is in
general less error-prone and already proved to yield high-quality
results as, for example, DBpedia itself is extracted primarily from
Wikipedia infoboxes; further approaches use the category system
of Wikipedia [
] or its list pages [
]. Many more
arXiv:2102.05444v1 [cs.IR] 10 Feb 2021
WWW ’21, April 19–23, 2021, Ljubljana, Slovenia N. Heist and H. Paulheim
approaches focus on tables (in Wikipedia or the web) as semi-
structured data source to extract entities and relations (see [
for a comprehensive survey). The focus of recent web table-based
approaches like Zhang et al. [
] is set on recognizing entities and
relationships within a table. Considering Fig. 1, the table below the
section Solo albums may be used to discover the publication years
of albums (relation extraction) or discover additional unknown al-
bums that are listed in further rows below Rubber and Swag (entity
and type detection).
The focus of this paper is broader with respect to two dimensions:
First, we extract information from any kind of structure where
similar entities co-occur. In Fig. 1, we would consider both tables
and lists (e.g. the list in the section Albums with Guns N’ Roses).
We refer to these co-occurrence structures as listings. Second, we
consider only the subject entities (SE) of listings. In our previous
work we dened SE with respect to Wikipedia list pages as "the
instances of the concept expressed by the list page" [
]. Considering
the List of Japanese speculative ction writers, its SE comprises all
Japanese speculative ction writers mentioned in listings of the
page. While in [
] the concept of SE is made explicit by the list
page, we deal with arbitrary listings in this paper. We thus assume
the concept may not be explicit or it may be indicated as part of
the page in which the listing appears (e.g. in the table header, or
the page title). Therefore, to each entity in a listing appearing as
instance to a common concept, we will further refer as subject entity.
The purpose of this work is to exploit the relationship between the
SE of a listing and the listing context. For Fig. 1, this means we
extract that all SE on the page’s listings are albums with the artist
Gilby Clarke, that The Spaghetti Incident? is an album by Guns N’
Roses, and so on.
To that end, we propose to learn these characteristics of a listing
with respect to the types and contextual relations of its SE. In
an ideal setting we know the SE of a listing and we are able to
retrieve all information about them from a knowledge graph – the
characteristics of a listing are then simply the types and relations
that are shared by all SE. But uncertainty is introduced by several
SE can only be determined heuristically. In previous work
], we achieved a precision of 90% for the recognition of
SE in Wikipedia listings.
Cross-domain knowledge graphs are not complete. Accord-
ing to the open world assumption (OWA), the absence of a
fact in a knowledge graph does not imply its incorrectness.
Web tables have a median of 6 rows,
and Wikipedia listings
have a median of 8 rows. Consequently, many listings only
have a small number of SE from which the characteristics
can be inferred.
As a result, considering each listing in isolation either leads
to a substantial loss of information (as listings with insucient
background information are disregarded) or to a high generaliza-
tion error (as decisions are made based on insucient background
We observe that the context of a listing is often a strong in-
dicator for its characteristics. In Fig. 1, the title of the top section
According to the WDC Web Table Corpus 2015:
Discography indicates that its listings contain some kind of musical
works, and the section title Albums with Guns N’ Roses provides
more detailed information. Our second observation is that these pat-
terns repeat when looking at a coherent data corpus. The Wikipedia
page of Axl Rose,
for example, contains the same constellation of
Considering listing characteristics with respect to their context
can thus yield in more general insights than considering every
listing in isolation. For example, the musical works of many artists
in Wikipedia are listed under the top section Discography. Hence,
we could learn the axioms
𝑡𝑜𝑝 𝑆𝑒𝑐𝑡𝑖𝑜𝑛.{"Discography"} ⊑ MusicalWork (1)
𝑡𝑜𝑝 𝑆𝑒𝑐𝑡𝑖𝑜𝑛 .{"Discography"} ⊑ ∃𝑎𝑟𝑡 𝑖𝑠𝑡 .{<𝑃 𝑎𝑔𝑒𝐸𝑛𝑡 𝑖𝑡𝑦>}(2)
which are then applicable to any listing with the top section Disco-
graphy in Wikipedia.
1.2 Approach and Contributions
In this work, we frame the task of nding descriptive rules for
listings based on their context as association rule mining problem
]. We dene rule metrics that take the inherent uncertainty into
account and make sure that rules are frequent (rule support), correct
(rule condence), and consistent over all listings (rule consistency).
Furthermore, we present an approach that executes the complete
pipeline from identication of SE to the extraction of novel entities
and assertions with Wikipedia as data corpus. To nd a reasonable
balance between correctness and coverage of the rules, we set
the thresholds based on a heuristic that takes the distribution of
named entity tags over entities as well as existing knowledge in a
knowledge graph into account. Applying the approach, we show
that we can enhance the knowledge graphs DBpedia with up to
2.9M entities and 8.3M assertions, and CaLiGraph
with up to 3M
entities and 30.4M assertions with an overall correctness of more
than 90%.
To summarize, the contributions of this paper are as follows:
We formulate the task of information extraction from co-
occurring similar entities in listings and show how to derive
descriptive rules for listing characteristics based on the list-
ing context (Sec. 3).
We present an approach that learns descriptive rules for
listings in Wikipedia and is capable of extracting several
millions of novel entities and assertions for Wikipedia-based
knowledge graphs (Sec. 4).
In our evaluation we demonstrate the high quality of the
extracted information and analyze the shortcomings of the
approach (Sec. 5).
The produced code is part of the CaLiGraph extraction frame-
work and publicly available.5
Information Extraction From Co-Occurring Similar Entities WWW ’21, April 19–23, 2021, Ljubljana, Slovenia
The work presented in this paper is a avour of knowledge graph
completion, more precisely, of adding new entities to a knowledge
graph [
]. We use rules based on page context to infer facts about
co-occurring entities. In particular, we focus on co-occurrence of
entities within document listings, where co-occurrence refers to
proximity w.r.t. page layout. Hence, in this section, we discuss
related works w.r.t. knowledge graph completion from listings,
exploitation of listing context, as well as rule learning for knowledge
2.1 Knowledge Graph Completion from
Knowledge graph completion using information in web tables has
been an already an active area of research in the last several years.
In 2016, Ritze et al. [
] proled the potential of web tables in the
WDC Web Table Corpus. Using the T2K Match framework, they
match web tables to DBpedia and nd that the best results for the
extraction of new facts can be achieved using knowledge-based
trust [
] (i.e., judging the quality of a set of extracted triples by
their overlap with the knowledge base). Zhang et al. [
] present
an approach for detection of novel entities in tables. They rst
exploit lexical and semantic similarity for entity linking and column
heading property matching. In a second step they use the output
to detect novel entities in table columns. Oulabi and Bizer [
tackle the same problem for Wikipedia tables with a bootstrapping
approach based on expert-dened rules. Macdonald and Barbosa
] extract new facts from Wikipedia tables to extend the Freebase
knowledge base. With an LSTM that uses contextual information
of the table, they extract new facts for 28 relations.
Lists have only very sparsely been used for knowledge graph
completion. Paulheim and Ponzetto [
] frame the general potential
of listings as a source of knowledge in Wikipedia. They propose
to use a combination of statistical and NLP methods to extract
knowledge and show that, by applying them to a single list page,
they are able to extract a thousand new statements.
Compared to all previously mentioned approaches, we take an
abstract view on listings by considering only their subject entities.
This provides the advantage that rules can be learned from and
applied to arbitrary listings. In addition to that, we do not only
discover novel entities, but also discover relations between those
novel entities and the page subject.
In our previous work [
], we have already presented an ap-
proach for the identication of novel entities and the extraction
of facts in Wikipedia list pages. List pages are pages in Wikipedia
that start with List of and contain listings (i.e., tables or lists) of
entities for a given topic (e.g. List of Japanese speculative ction
writers). The approach is divided into two phases: In a rst phase, a
dataset of tagged entities from list pages is extracted. With distant
supervision from CaLiGraph, a knowledge graph with a detailed
type hierarchy derived from Wikipedia categories and list pages,
a part of the mentioned entities is heuristically labeled as subject
entities and non-subject entities. In a second phase, the dataset is
enriched with positional, lexical, and statistical features extracted
from the list pages. On the basis of this data, an XGBoost classi-
er is able to identify more than two million subject entities with
an average precision of 90%. As not all the information about the
subject entities is contained in the knowledge graphs DBpedia and
CaLiGraph, they can be enhanced with the missing information.
In this work, we reuse the approach presented in [
] for identi-
fying subject entities. Further, as it is the only approach that also
works with arbitrary listings, we use it as a baseline in our ex-
periments. As, in its current state, it only works for list pages in
Wikipedia, we extend it to arbitrary pages with a simple frequency-
based approach.
2.2 Exploiting the Context of Listings
As tables are the more actively researched type of listings, we focus
here on the types of context used when working with tables. The
most obvious source of context is found directly on the page where
the table is located. This page context is, for example, used by
InfoGather [
] to detect possible synonyms in table headers for
means of table matching.
Zhang [
] distinguishes between "in-table" features like the
table header, and "out-table" features like captions, page title, and
text of surrounding paragraphs. With both kinds of features, they
perform entity disambiguation against Freebase.
The previously mentioned approach of Macdonald and Barbosa
] focuses on tables in Wikipedia and hence uses specic context
features like section titles, table headers and captions, and the text
in the rst paragraph of the table’s section. Interestingly, they do
not only discover relations between entities in the table, but also
between a table entity and the page subject.
] leverages patterns occurring in headers of Wikipe-
dia tables to consistently discover DBpedia relations. Lehmberg et
al. [
] tackle the problem of small web tables with table stitching,
i.e., they combine several small tables with a similar context (e.g.,
same page or domain and matching schema) to a larger one, making
it easier to extract facts from it.
Apart from page context, many approaches use the context of
entities in tables to improve extraction results. Zhang et al. [
generate new sub-classes to a taxonomy for a set of entities. There-
fore, they nd the best-describing class using the context of the
entities. In particular, they use the categories of the entities as well
as the immediate context around the entities on the page. Another
approach that uses entity categories as context is TableNet [
]. They
leverage the context to nd schematically similar or related tables
for a given table in Wikipedia.
In our experiments with Wikipedia, we use section headers as
page context and types in the knowledge graph as entity context.
However, the denition of context in our approach is kept very gen-
eric on purpose. By doing that, we are able to incorporate additional
context sources like section text or entity categories to improve
extraction results. This, however, also comes with an increase in
rule complexity and, consequently, run time.
2.3 Rule-based Knowledge Graph Completion
Rule-based knowledge graph completion approaches typically gen-
erate rules either on instance-level (rules that add new facts for
individual instances) or on schema-level (rules that add additional
schematic constraints).
WWW ’21, April 19–23, 2021, Ljubljana, Slovenia N. Heist and H. Paulheim
] and AnyBURL [
] are instance-level rule learners
inspired by integer linear programming (ILP). The former uses top-
down, the latter bottom-up rule learning to generate rules in the
fashion of 𝑏𝑜𝑟𝑛 (𝑋 , 𝐴) 𝑐𝑎𝑝𝑖 𝑡𝑎𝑙 (𝐴, 𝑌 )=𝑐𝑖𝑡𝑖𝑧𝑒𝑛 (𝑋 , 𝑌 ).
DL-Learner [
] is an ILP-based approach on schema-level which
nds description logic patterns for a set of instances. A related
approach uses statistical schema induction [
] to derive additional
schema constraints (e.g. range restrictions for predicates).
The above mentioned approaches are merely link prediction ap-
proaches, i.e. they predict new relations between entities already
contained in the knowledge graph. The same holds for the om-
nipresent knowledge graph embedding approaches [
]. Such ap-
proaches are very productive when enough training data is available
and they provide exact results especially when both positive and
negative examples are given. In the setting of this paper, we are
working with (more or less) noisy external data.
With regard to instance- versus schema-level, our approach can
be regarded as a hybrid approach that generates rules for sets of
entities, which are in turn used to generate facts on an instance-
level. In this respect, our approach is similar to C-DF [
] which
uses Wikipedia categories as an external data source to derive the
characteristics of categories. To that end, they derive lexical patterns
from category names and contained entities.
In this paper, we apply rule learning to co-occurring entities in
Wikipedia. While existing approaches have only considered explicit
co-occurrence, i.e., categories or list pages, we go beyond the state
of the art by considering arbitrary listings in Wikipedia, as the one
shown in Fig. 1.
In this paper, we consider a data corpus
from which co-occurring
entities can be extracted (e.g., listings in Wikipedia or a collection
of spreadsheets). Furthermore, we assume that a knowledge graph
which contains a subset of those entities can be extended with
information learned about those co-occurring entities.
3.1 Task Formulation
The Knowledge Graph
is a set of assertions about its entities in
the form of triples
{(𝑠 , 𝑝, 𝑜 )|𝑠∈ E, 𝑝 ∈ P , 𝑜 ∈ E ∪ T ∪ L}
over sets of entities
, predicates
, types
, and literals
. We
refer to statements about the types of an entity (i.e.,
𝑝=rdf:type, 𝑜
) as type assertions (
), and to statements about relations
between two entities (i.e.,
𝑜∈ E
) as relation assertions (
K⊇ K
, we refer to the idealized complete version of
With regard to the OWA this means that a fact is incorrect if it is
not contained in K.6
The data corpus
contains a set of listings
, where each listing
contains a number of subject entities
. Our task is to
identify statements that hold for all subject entities
in a listing
. We distinguish taxonomic and relational information that is
expressed in K.
is merely a theoretical construct, since a complete knowledge graph of all entities
in the world cannot exist.
The taxonomic information is a set of types that is shared by all
SE of a listing:
𝜙={𝑡|𝑡 T ,𝑠𝑆𝐸𝜙:(𝑠, rdf:type, 𝑡 ) ∈ K},(3)
and the relational information is a set of relations to other entities
which is shared by all SE of a listing:7
R𝜙={(𝑝, 𝑜 )|𝑝 P ∪ P 1, 𝑜 ∈ E,𝑠𝑆𝐸𝜙:(𝑠 , 𝑝, 𝑜 ) ∈ K}.(4)
From these characteristics of listings, we can derive all the addi-
tional type assertions
{(𝑠 , rdf:type, 𝑡 )|𝑠𝑆𝐸𝜙, 𝑡 ∈ T
𝜙} \ 𝑇 𝐴 (5)
and additional relation assertions
{(𝑠 , 𝑝, 𝑜 )|𝑠𝑆𝐸𝜙,(𝑝 , 𝑜) ∈ R𝜙} \ 𝑅𝐴 (6)
that are encoded in
and missing in
. Furthermore,
can contain additional entities that are not yet contained in
as there is no restriction for subject entities of Φto be part of K.
For the sake of readability, we will only describe the case of
for the remainder of this section as
is – notation-wise – a special
case of R𝜙with 𝑝=rdf:type and 𝑜 T .
3.2 Learning Descriptive Rules for Listings
Due to the incompleteness of
, it is not possible to derive the
exact set of relations
for every listing in
. Hence, our goal is to
derive an approximate version
by using
and the knowledge
about 𝑆𝐸𝜙in K.
Similar to the rule learner AMIE+ [
], we use the partial com-
pleteness assumption (PCA) to generate negative evidence. The
PCA implies that if
(𝑠, 𝑝 , 𝑜) ∈ K
(𝑠, 𝑝 , 𝑜 ) ∈ K=
(𝑠, 𝑝 , 𝑜 ) ∈ K
. In order words, if
makes some assertions with a
for a subject
, then we assume that
contains every
𝑝-related information about 𝑠.
Following from the PCA, we use the
of entities with a
specic predicate-object combination in a set of entities 𝐸
𝑐𝑜𝑢𝑛𝑡 (𝐸 , 𝑝, 𝑜 )=|{𝑠|𝑠𝐸, 𝑜:(𝑠, 𝑝 , 𝑜) ∈ K}| (7)
and the
of entities having predicate
with an arbitrary
𝑐𝑜𝑢𝑛𝑡 (𝐸 , 𝑝)=|{𝑠|𝑠𝐸, 𝑜:(𝑠, 𝑝 , 𝑜) ∈ K}| (8)
to compute a maximum-likelihood-based frequency of a specic
predicate-object combination occurring in 𝐸:
𝑓 𝑟𝑒 𝑞(𝐸, 𝑝 , 𝑜)=
𝑐𝑜𝑢𝑛𝑡 (𝐸 , 𝑝, 𝑜 )
𝑐𝑜𝑢𝑛𝑡 (𝐸 , 𝑝).(9)
From Eq. 9 we rst derive a naive approximation of a listing’s
relations by including all relations with a frequency above a dened
threshold 𝜏𝑓 𝑟𝑒𝑞 :
R𝑓 𝑟𝑒 𝑞
={(𝑝, 𝑜 )|(𝑝, 𝑜) ∈ R, 𝑓 𝑟 𝑒𝑞 (𝑆𝐸𝜙, 𝑝, 𝑜 )>𝜏𝑓 𝑟 𝑒𝑞 }.(10)
Here, the entities in
may occur both in the subject as well as in the object
position. But for a more concise notation, we use only (p,o)-tuples and introduce the
set of inverse predicates
to express that SE may also occur in object position.
This is, however, only a notation and the inverse predicates do not have to exist in the
Information Extraction From Co-Occurring Similar Entities WWW ’21, April 19–23, 2021, Ljubljana, Slovenia
Table 1: Exemplary context (𝜁), type frequency (𝑇𝐹), and re-
lation frequency (𝑅𝐹) vectors for a set of listings extracted
from 𝐷. While 𝜁is extracted directly from 𝐷,𝑇𝐹and 𝑅𝐹are
retrieved via distant supervision from K.
Listing 𝜁 𝑇 𝐹𝑅𝐹
𝜙1(1 0 1 ... 1) (0.2 0.9 0.0 ... 0.1) (0.9 0.1 0.0 ... 0.1)
𝜙2(0 1 1 ... 0) (0.0 0.2 0.0 ... 0.9) (0.0 0.0 0.0 ... 0.2)
𝜙3(0 0 0 ... 0) (0.7 0.7 0.0 ... 0.0) (0.0 0.0 0.0 ... 0.4)
𝜙𝑛1(1 0 0 ... 1) (0.8 0.9 0.0 ... 0.0) (0.0 0.9 0.0 ... 0.0)
𝜙𝑛(1 0 0 ... 1) (0.7 1.0 0.0 ... 0.3) (0.0 0.0 0.8 ... 0.0)
As argued in Sec. 1.1, we improve this naive frequency-based
approximation by learning more general patterns that describe the
characteristics of listings using their context.
Hypothesis 1.
The context
of a listing
contains such
information about
that it can be used to nd subsets of
similar R.
Let Table 1 contain the information about all listings in
. A
is dened by its context
(which can in theory contain
any information about
, from the title of its section to an actual
image of the listing), the type frequencies
(𝑡1, 𝑡2, .., 𝑡𝑥) 𝑇𝐹
, and
the relation frequencies
(𝑟1, 𝑟2, . .,𝑟 𝑦) ∈ 𝑅𝐹
. Listings
, and
have overlapping context vectors.
has a consistently high
frequency over all three listings. It is thus a potential type charac-
teristic for this kind of listing context. Furthermore,
has a high
frequency in
, and
– if the three relations
share the same predicate, they may all express a similar relation to
an entity in their context (e.g. to the entity the page is about).
In a concrete scenario, the context vector (1 0 0 ... 1) might
indicate that the listing is located on the page of a musician under
the section Solo albums.
holds the frequency of the type Album
in this listing and
describe the frequencies of the relations
(artist, Gilby Clarke), (artist, Axl Rose), and (artist, Slash).
We formulate the task of discovering frequent co-occurrences
of context elements and taxonomic and relational patterns as an
association rule mining task over all listings in
. Association
rules, as introduced by Agrawal et al. [
], are simple implication
patterns originally developed for large and sparse datasets like
transaction databases of supermarket chains. To discover items that
are frequently bought together, rules of the form
produced, with
being itemsets. In the knowledge graph
context, they have been used, e.g., for enriching the schema of a
knowledge graph [23, 30].
For our scenario, we need a mapping from a context vector
to a predicate-object tuple. Hence, we dene a rule
, its antecedent
𝑟𝑎, and its consequent 𝑟𝑐as follows:
𝑟:𝑟𝑎𝑍=𝑟𝑐∈ ( P ∪ P 1) × (T ∪ E ∪ X).(11)
As a rule should be able to imply relations to entities that vary with
the context of a listing (e.g. to Gilby Clarke as the page’s main entity
in Fig. 1), we introduce
as the set of placeholders for specic
context entities (instead of Gilby Clarke, the object of the rule’s
consequent would be <PageEntity>).
We say a rule antecedent
matches a listing context
if the vector of
is 1 when the vector of
is 1. In essence,
must comprise
. Accordingly, we need to nd a set of rules
, so
that for every listing 𝜙the set of approximate listing relations
R𝑟𝑢𝑙 𝑒
resembles the true relations R𝜙as closely as possible.
Considering all the listings in Fig. 1, their
R𝑟𝑢𝑙 𝑒
should, among
others, contain the rules8,9
𝑡𝑜𝑝𝑆𝑒𝑐𝑡 𝑖𝑜𝑛("Discography")=⇒ (𝑡𝑦𝑝𝑒 , MusicalWork)(13)
𝑡𝑜𝑝𝑆𝑒𝑐𝑡 𝑖𝑜𝑛("Discography")=⇒ (𝑎𝑟𝑡𝑖 𝑠𝑡, <𝑃𝑎𝑔𝑒𝐸𝑛𝑡𝑖𝑡𝑦 >).(14)
It is important to note that these rules can be derived from
listings with diering context vectors. All listings only have to have
in common that their top section has the title Discography and that
the contained entities are mostly of
with the page
entity as artist. Still, the individual listings may, for example, occur
in sections with dierent titles.
3.3 Quality Metrics for Rules
In original association rule mining, two metrics are typically con-
sidered to judge the quality of a rule
: the support of the
rule antecedent (how often does
occur in the dataset), and the
condence of the rule (how often does
occur in relation to
Transferring the support metric to our task, we count the ab-
solute frequency of a particular context occurring in
. Let
{𝜙|𝜙Φ, 𝑟𝑎𝜁𝜙}
, then we dene the support of the rule ante-
cedent 𝑟𝑎as
𝑠𝑢𝑝 𝑝 (𝑟𝑎)=|Φ𝑟𝑎|.(15)
Due to the incompleteness of
, the values of
are in our case
no denitive items but maximum-likelihood estimates of types and
relations. With respect to these estimates, a good rule has to fulll
two criteria: it has to be correct (i.e. frequent with respect to all SE
of the covered listings) and it has to be consistent (i.e. consistently
correct over all the covered listings).
We dene the correctness, or condence, of a rule as the fre-
quency of the rule consequent over all SE of a rule’s covered list-
𝑐𝑜𝑛 𝑓 (𝑟)=Í𝜙Φ𝑟𝑎𝑐𝑜𝑢𝑛𝑡 (𝑆 𝐸𝜙, 𝑝𝑟𝑐, 𝑜 𝑟𝑐)
Í𝜙Φ𝑟𝑎𝑐𝑜𝑢𝑛𝑡 (𝑆𝐸𝜙, 𝑝𝑟𝑐),(16)
and we dene the consistency of a rule using the mean abso-
lute deviation of an individual listing’s condence to the overall
condence of the rule:
𝑐𝑜𝑛𝑠 (𝑟)=1Í𝜙Φ𝑟𝑎|𝑓 𝑟𝑒 𝑞(𝑆𝐸𝜙, 𝑝𝑟𝑐, 𝑜𝑟𝑐) 𝑐𝑜𝑛 𝑓 (𝑟) |
𝑠𝑢𝑝 𝑝 (𝑟𝑎).(17)
While a high condence ensures that the overall assertions gen-
erated by the rule are correct, a high consistency ensures that few
listings with many SE do not outvote the remaining covered listings.
Note that Eqs. 1 and 2 are the axiom equivalents of Eqs. 13 and 14. For better readability,
we use the description logics notation of Eqs. 1 and 2 from here on.
9Instead of a binary vector, we use a more expressive notation for the listing context
in our examples. The notations are trivially convertible by one-hot-encoding.
WWW ’21, April 19–23, 2021, Ljubljana, Slovenia N. Heist and H. Paulheim
To select an appropriate set of rules
from all the candidate
in the search space, we have to pick reasonable thresholds
for the minimum support (
), the minimum condence (
𝜏𝑐𝑜𝑛 𝑓
and the minimum consistency (
). By applying these thresholds,
we nd our nal set of descriptive rules 𝑅:
{𝑟|𝑟𝑅, 𝑠𝑢𝑝 𝑝 (𝑟𝑎)>𝜏𝑠𝑢𝑝𝑝 𝑐𝑜𝑛𝑓 (𝑟)>𝜏𝑐𝑜𝑛 𝑓 𝑐𝑜 𝑛𝑠 (𝑟)>𝜏𝑐𝑜𝑛𝑠 }.
Typically, the choice of these thresholds is strongly inuenced by
the nature of the dataset
and the extraction goal (correctness
versus coverage).
Wikipedia is a rich source of listings, both in dedicated list pages
as well as in sections of article pages. Hence, we use it as a data
corpus for our experiments. In Sec. 6, we discuss other appropriate
corpora for our approach.
Due to its structured and encyclopedic nature, Wikipedia is a
perfect application scenario for our approach. We can exploit the
structure by building very expressive context vectors. Obviously,
this positively inuences the quality of extraction results. Still, the
denition of the context vector is kept abstract on purpose to make
the approach applicable to other kinds of web resource as well.
However, an empirical evaluation of the practicability or perform-
ance of the approach for resources outside of the encyclopedic
domain is out of scope of this paper.
4.1 Approach Overview
Fig. 2 gives an overview of our extraction approach. The input
of the approach is a dump of Wikipedia as well as an associated
knowledge graph. In the Subject Entity Discovery phase, listings and
their context are extracted from the Wikipedia dump and subject
entities are identied (Sec. 4.3). Subsequently, the existing informa-
tion in the knowledge graph is used to mine descriptive rules from
the extracted listings (Sec. 4.4). Finally, the rules are applied to all
the listings in Wikipedia in order to extract new type and relation
assertions (Sec. 4.5).
4.2 Wikipedia as a Data Corpus
We pick Wikipedia as a data corpus for our experiments as it brings
several advantages:
Structure. Wikipedia is written in an entity-centric style with a
focus on facts. Listings are often used to provide an overview of
a set of entities that are related to the main entity. Due to the en-
cyclopedic style and the peer-reviewing process, it has a consistent
structure. Especially section titles are used consistently for specic
topics. Wikipedia has its own markup language (Wiki markup),
which allows a more consistent access to interesting page struc-
tures like listings and tables than plain HTML.
Entity Links. If a Wikipedia article is mentioned in another art-
icle, it is typically linked in the Wiki markup (a so called blue link).
Furthermore, it is possible to link to an article that does not (yet)
exist (a so called red link). As Wikipedia articles can be trivially
mapped to entities in Wikipedia-based knowledge graphs like DB-
pedia, since they create one entity per article, we can identify many
named entities in listings and their context without the help of an
entity linker.
For our experiments, we use a Wikipedia dump of October 2016
which is, at the time of the experiments, the most recent dump that
is compatible with both DBpedia and CaLiGraph. In this version,
Wikipedia contains 6.9M articles, 2.4M of which contain listings
with at least two rows.
In total, there are 5.1M listings with a row
count median of 8, mean of 21.9, and standard deviation of 76.8. Of
these listings, 1.1M are tables, and 4.0M are enumerations.
4.3 Subject Entity Discovery
4.3.1 Entity Tagging. Apart from the already tagged entities via
blue and red links, we have to make sure that any other named
entity in listings and their context is identied as well. This is done
in two steps:
In a rst step, we expand all the blue and red links in an article. If
a piece of text is linked to another article, we make sure that every
occurrence of that piece of text in the article is linked to the other
article. This is necessary as by convention other articles are only
linked at their rst occurrence in the text.11
In a second step, we use a named entity tagger to identify ad-
ditional named entities in listings. To that end, we use a state-of-
the-art entity tagger from spaCy.
This tagger is trained on the
corpus, and thus not specically trained to identify
named entities in short text snippets like they occur in listings.
Therefore, we specialize the tagger by providing it Wikipedia list-
ings as additional training data with blue links as positive examples.
In detail, the tagger is specialized as follows:
We retrieve all listings in Wikipedia list pages as training
We apply the plain spaCy entity tagger to the listings to get
named entity tags for all mentioned entities.
To make these tags more consistent, we use information
from DBpedia about the tagged entities: We look at the dis-
tribution of named entity tags over entities with respect to
their DBpedia types and take the majority vote. For example,
if 80% of entities with the DBpedia type
are annot-
ated with the tag PERSON, we use PERSON as label for all
these entities.
Using these consistent named entity tags for blue-link entit-
ies, we specialize the spaCy tagger.
4.3.2 Subject Entity Classification. We apply the approach from
] for the identication of subject entities in listings. In short, we
use lexical, positional, and statistical features to classify entities as
subject or non-subject entities (refer to Sec. 2.1 for more details).
Despite being developed only for listings in list pages, the classier
is applicable to any kind of listing in Wikipedia. A disadvantage
of this broader application is that the classier is not trained in
such a way that it ignores listings used for organisational or design
purposes (e.g. summaries or timelines). These have to be ltered
out in the subsequent stages.
Wiki markup is parsed with WikiTextParser:
Information Extraction From Co-Occurring Similar Entities WWW ’21, April 19–23, 2021, Ljubljana, Slovenia
Figure 2: An overview of the approach with exemplary outputs of the individual phases.
4.3.3 Results. After expanding all the blue and red links on the
pages, the dataset contains 5.1M listings with 60.1M entity mentions.
51.6M additional entity mentions are identied by the named entity
Of all the entity mentions, we classify 25.8M as subject entities.
Those occur in 2.5M listings of 1.3M pages. This results in a mean
of 10.5 and median of 4 subject entities per listing with a standard
deviation of 49.8.
4.4 Descriptive Rule Mining
4.4.1 Describing Listings. The search space for rule candidates is
dened by the listing context. Thus, we choose the context in such
a way that it is expressive enough to be an appropriate indicator
, and concise enough to explore the complete search
space without any additional heuristics.
We exploit the fact that Wikipedia pages of a certain type (e.g.,
musicians) mostly follow naming conventions for the sections of
their articles (e.g., albums and songs are listed under the top sec-
tion Discography). Further, we exploit that the objects of the SE’s
relations are usually either the entity of the page, or an entity men-
tioned in a section title. We call these typical places for objects the
relation targets. In Fig. 1, Gilby Clarke is an example of a PageEntity
target, and Guns N’ Roses as well as Nancy Sinatra are examples for
SectionEntity targets. As a result, we use the type of the page entity,
the top section title, and the section title as listing context.
Additionally, we allow to use the type of an entity mentioned
in a section instead of the section title itself. This provides more
exibility by learning abstract rules, e.g., to distinguish between
albums (listed in a section describing a band):
𝑝𝑎𝑔𝑒𝐸𝑛𝑡 𝑖𝑡𝑦𝑇𝑦𝑝𝑒 .{Person } ⊓ ∃𝑡𝑜 𝑝𝑆 𝑒𝑐𝑡𝑖𝑜𝑛.{"Discography"}
⊓∃𝑠𝑒𝑐𝑡 𝑖𝑜𝑛𝐸𝑛𝑡𝑖𝑡𝑦𝑇 𝑦𝑝𝑒.{Band} Album,
and songs (listed in a section describing an album):
𝑝𝑎𝑔𝑒𝐸𝑛𝑡 𝑖𝑡𝑦𝑇𝑦𝑝𝑒 .{Person } ⊓ ∃𝑡𝑜 𝑝𝑆 𝑒𝑐𝑡𝑖𝑜𝑛.{"Discography"}
⊓∃𝑠𝑒𝑐𝑡 𝑖𝑜𝑛𝐸𝑛𝑡𝑖𝑡𝑦𝑇 𝑦𝑝𝑒.{Album} Song.
4.4.2 Threshold Selection. We want to pick the thresholds in such
a way that we tolerate some errors and missing information in
but do not allow many over-generalized rules that create incorrect
assertions. Our idea for a sensible threshold selection is based on
two assumptions:
Assumption 1.
Being based on a maximum-likelihood estima-
tion, rule condence and consistency roughly order rules by the
degree of prior knowledge we have about them.
Assumption 2.
Assertions generated by over-generalized rules
contain substantially more random noise than assertions generated
by good rules.
Assumption 1 implies that the number of over-generalized rules
increases with the decrease of condence and consistency. As a
consequence, assumption 2 implies that the amount of random
noise increases with decrease of condence and consistency.
To measure the increase of noise in generated assertions, we
implicitly rely on existing knowledge in
by using the named
entity tags of subject entities as a proxy. This works as follows:
For a subject entity
that is contained in
, we have its type
and we have its named entity tag
our named entity tagger. Going over all SE of listings in
, we
compute the probability of an entity with type
having the tag
by counting how often they co-occur:
WWW ’21, April 19–23, 2021, Ljubljana, Slovenia N. Heist and H. Paulheim
𝑡𝑎𝑔𝑝𝑟 𝑜𝑏 (𝑡, 𝜓 )=
|{𝑒|𝜙Φ:𝑒𝑆𝐸𝜙𝑡∈ T
|{𝑒|𝜙Φ:𝑒𝑆𝐸𝜙𝑡∈ T
𝑒}| .(19)
For example, for the DBpedia type
, we nd the tag prob-
WORK_OF_ART: 0.49, ORG: 0.14, PRODUCT : 0.13, PERSON : 0.07,
showing that album titles are rather dicult to recognize. For the
and the tag PERSON, on the other hand, we nd a
probability of 0.86.
We can then compute the tag-based probability for a set of asser-
by averaging over the tag probability that is produced by
the individual assertions. To compute the latter, we compare the tag
of the assertion’s subject entity with some kind of type information
about it. This type information is either the asserted type (in case
of a type assertion), or the domain of the predicate
(in case of a
relation assertion):
𝑡𝑎𝑔 𝑓 𝑖𝑡 (𝐴)=
Í(𝑠,𝑝,𝑜 )∈𝐴𝑡 𝑎𝑔𝑝𝑟 𝑜𝑏 (𝑜,𝜓𝑠)
|𝐴|if 𝑝=rdf:type
Í(𝑠,𝑝,𝑜 )∈𝐴𝑡 𝑎𝑔𝑝𝑟 𝑜𝑏 (𝑑𝑜𝑚𝑎𝑖𝑛𝑝,𝜓𝑠)
While we do not expect the named entity tags to be perfect, our
approach is based on the idea that the tags are consistent to a large
extent. By comparing the
𝑡𝑎𝑔 𝑓 𝑖𝑡
of assertions produced by rules
with varying levels of condence and consistency, we expect to see
a clear decline as soon as too many noisy assertions are added.
4.4.3 Results. Fig. 3 shows the
𝑡𝑎𝑔 𝑓 𝑖𝑡
for type and relation asser-
tions generated with varying levels of rule condence and consist-
ency. Our selection of thresholds is indicated by blue bars, i.e. we
set the thresholds to the points where the
𝑡𝑎𝑔 𝑓 𝑖𝑡
has its steepest
drop. The thresholds are picked conservatively to select only high-
quality rules by selecting points before an accelerated decrease of
𝑡𝑎𝑔 𝑓 𝑖𝑡
. But more coverage-oriented selections are also
possible. In Fig. 3d, for example, a threshold of 0.75 is also a valid
An analysis of rules with dierent levels of condence and con-
sistency has shown that a minimum support for types is not ne-
cessary. For relations, a support threshold of 2 is helpful to discard
over-generalized rules. Further, we found that it is acceptable to
pick the thresholds independently from each other, as the turning
points for a given metric don’t vary signicantly when varying the
remaining metrics.
Applying these thresholds, we nd an overall number of 5,294,921
type rules with 369,139 distinct contexts and 244,642 distinct types.
Further, we nd 3,028 relation rules with 2,602 distinct contexts
and 516 distinct relations. 949 of the relation rules have the page
entity as target, and 2,079 have a section entity as target.
Among those rules are straightforward ones like
𝑝𝑎𝑔𝑒𝐸𝑛𝑡 𝑖𝑡𝑦𝑇𝑦𝑝𝑒 .{Person } ⊓ ∃𝑡𝑜 𝑝𝑆 𝑒𝑐𝑡𝑖𝑜𝑛.{"Acting lmography"}
⊑ ∃𝑎𝑐𝑡𝑜 𝑟 .{<𝑃𝑎𝑔𝑒𝐸𝑛𝑡𝑖𝑡𝑦>},
We use the domain of the predicate
as dened in
. In case of
𝑝∈ P1
, we use
the range of the original predicate.
and more specic ones like
𝑝𝑎𝑔𝑒𝐸𝑛𝑡 𝑖𝑡𝑦𝑇𝑦𝑝𝑒 .{Location } ⊓ ∃𝑡𝑜 𝑝𝑆 𝑒𝑐𝑡𝑖𝑜𝑛.{"Media"}
⊓∃𝑠𝑒𝑐𝑡 𝑖𝑜𝑛.{"Newspapers"} ⊑ Periodical_literature.
4.5 Assertion Generation and Filtering
4.5.1 Assertion Generation. We apply the rules selected in the
previous section to the complete dataset of listings to generate type
and relation assertions. Subsequently, we remove any duplicate
assertions and assertions that already exist in K.
4.5.2 Tag-based Filtering. To get rid of errors introduced during
the extraction process (e.g. due to incorrectly extracted subject
entities or incorrect rules), we employ a nal ltering step for the
generated assertions: every assertion producing a
𝑡𝑎𝑔𝑝𝑟 𝑜𝑏 1
discarded. The rationale behind the threshold is as follows: Types
have typically one and sometimes two corresponding named entity
tags (e.g. the tag PERSON for the DBpedia type
, or the tags
ORG and FAC for the type
). As tag probabilities are relative
frequencies, we make sure that, with a threshold of
, at most two
tags are accepted for any given type.
For the tag probabilities of type
from Sec. 4.4.2, the only
valid tag is WORK_OF_ART. As a consequence, any assertions
of the form
(𝑠, 𝑟 𝑑 𝑓 :𝑡𝑦 𝑝𝑒, Album)
having a tag other than
WORK_OF_ART are discarded.
4.5.3 Results. Tab. 2 shows the number of generated type and rela-
tion assertions before and after the tag-based ltering. The number
of inferred types are listed separately for DBpedia and CaLiGraph.
For relations, we show two kinds: The entry Relations lists the num-
ber of extracted assertions from rules. As DBpedia and CaLiGraph
share the same set of predicates, these assertions are applicable to
both graphs. Furthermore, as Relations (via CaLiGraph), we list the
number of relations that can be inferred from the extracted CaLi-
Graph types via restrictions in the CaLiGraph ontology. CaLiGraph
contains more than 100K of such restrictions that imply a relation
based on a certain type. For example, the ontology contains the
value restriction
Pop_rock_song ⊑ ∃𝑔𝑒𝑛𝑟𝑒 .{Pop music}.
As we extract the type
for the Beach Boys song At
My Window, we infer the fact (At My Window, 𝑔𝑒𝑛𝑟𝑒 , Pop music).
For CaLiGraph, we nd assertions for 3.5M distinct subject en-
tities with 3M of them not contained in the graph. For DBpedia,
we nd assertions for 3.1M distinct subject entities with 2.9M of
them not contained. The unknown subject entities are, however,
not disambiguated yet. Having only small text snippets in listings
as information about these entities, a disambiguation with general-
purpose disambiguation approaches [
] is not practical. We thus
leave this as an own research topic for future work. For an estim-
ation of the actual number of novel entities, we rely on previous
work [
], where we analyzed the overlap for red links in list pages.
In that paper, we estimate an overlap factor of 1.07 which would –
when applied to our scenario – reduce the number of actual novel
entities to roughly 2.8M for CaLiGraph and 2.7M for DBpedia. In re-
lation to the current size of those graphs, this would be an increase
of up to 38% and 54%, respectively [9].
Information Extraction From Co-Occurring Similar Entities WWW ’21, April 19–23, 2021, Ljubljana, Slovenia
(a) Type condence (b) Type consistency (c) Relation condence (d) Relation consistency
Figure 3: 𝑡 𝑎𝑔𝑓 𝑖 𝑡 of assertions generated from rules in a specied condence or consistency interval. Bars show scores for a
given interval (e.g. (0.75,0.80]), lines show cumulative scores (e.g. (0.75,1.00]). Blue bars indicate the selected threshold.
Table 2: Number of generated assertions after removing ex-
isting assertions (Raw), and after applying tag-based lter-
ing (Filtered).
Assertion Type Raw Filtered
Types (DBpedia) 11,459,047 7,721,039
Types (CaLiGraph) 47,249,624 29,128,677
Relations 732,820 542,018
Relations (via CaLiGraph) 1,381,075 796,910
Table 3: Correctness of manually evaluated assertions.
Assertion Type #Dataset #Samples Correct [%]
Types (DBpedia)
frequency-based 6,680,565 414 91.55 ±2.68
rule-based 7,721,039 507 93.69 ±2.12
Types (CaLiGraph)
frequency-based 26,676,191 2,000 89.40 ±1.23
rule-based 29,128,677 2,000 91.95 ±1.19
frequency-based 392,673 1,000 93.80 ±1.49
rule-based 542,018 1,000 95.90 ±1.23
In our performance evaluation, we judge the quality of generated
assertions from our rule-based approach. As a baseline, we ad-
ditionally evaluate assertions generated by the frequency-based
approach (see Eq. 10). For the latter, we use a threshold comparable
to our rule-based approach (i.e., we set
𝜏𝑓 𝑟𝑒 𝑞
𝜏𝑐𝑜𝑛 𝑓
and disregard
listings with less than three subject entities).
5.1 Evaluation Procedure
The evaluated assertions are created with a stratied random sampling
strategy. The assertions are thus distributed proportionally over all
page types (like
) and sampled randomly within
The labeling of the assertions is performed by the authors with
the procedure as follows: For a given assertion, rst the page of
the listing is inspected, then – if necessary and available – the
page of the subject entity. If a decision cannot be made based on
this information, a search engine is used to evaluate the assertion.
Samples of the rule-based and frequency-based approaches are
evaluated together and in random order to ensure objectivity.
Tab. 3 shows the results of the performance evaluation. In total,
we evaluated 2,000 examples per approach for types and 1,000
examples per approach for relations. The taxonomy of CaLiGraph
comprises the one of DBpedia. Thus, we evaluated the full sample
for CaLiGraph types and report the numbers for both graphs, which
is the reason why the sample size for DBpedia is lower. For relations,
we only evaluate the ones that are generated directly from rules
and not the ones inferred from CaLiGraph types, as the correctness
of the inferred relations directly depends on the correctness of
CaLiGraph types.
5.2 Type and Relation Extraction
The evaluation results in Tab. 3 show that the information extracted
from listings in Wikipedia is of an overall high quality. The rule-
based approach yields a larger number of assertions with a higher
correctness for both types and relations.
For both approaches, the correctness of the extracted assertions
is substantially higher for DBpedia. The reason for that lies in the
diering granularity of their taxonomies. DBpedia has 764 dierent
types while CaLiGraph has 755,441 with most of them being more
specic extensions of DBpedia types. For example, DBpedia might
describe a person as
, while CaLiGraph describes it as
. The average
depth of predicted types is 2.06 for the former and 3.32 for the
While the asserted types are very diverse (the most predicted
type is
with 7.5%), asserted relations are dominated by the
predicate genus with 69.8% followed by isPartOf (4.4%) and artist
(3.2%). This divergence cannot be explained with a dierent cover-
age: In DBpedia, 72% of entities with type
have a genus,
and 69% of entities with type
have an artist. But we
identify two other inuencing factors: Wikipedia has very specic
guidelines for editing species, especially with regard to standardiz-
ation and formatting rules.
In addition to that, the genus relation
is functional and hence trivially fullling the PCA. As our approach
is strongly relying on this assumption and it potentially inhibits
the mining of practical rules for non-functional predicates (like,
WWW ’21, April 19–23, 2021, Ljubljana, Slovenia N. Heist and H. Paulheim
for example, for artist), we plan on investigating this relationship
The inferred relations from CaLiGraph types are not evaluated
explicitly. However, based on the correctness of restrictions in
CaLiGraph that is reported to be 95.6% [
] and from the correctness
of type assertions, we estimate the correctness of the resulting
relation assertions to be around 85.5% for the frequency-based and
around 87.9% for the rule-based approach.
5.3 Novel Entity Discovery
For CaLiGraph, the frequency-based approach nds assertions for
2.5M distinct subject entities (2.1M of them novel). While the rule-
based approach nds 9% more assertions, its assertions are distrib-
uted over 40% more entities (and over 43% more novel entities).
This demonstrates the capabilities of the rule-based approach to ap-
ply contextual patterns to environments where information about
actual entities is sparse.
Further, we analyzed the portion of evaluated samples that ap-
plies to novel entities and found that the correctness of these state-
ments is slightly better (between 0.1% and 0.6%) than the overall
correctness. Including CaLiGraph types, we nd an average of 9.03
assertions per novel entity, with a median of 7. This is, again, due to
the very ne-grained type system of CaLiGraph. For example, for
the rapper Dizzle Don, which is a novel entity, we nd 8 types (from
) and 4 relations: (occu-
pation, Singing), (occupation, Rapping), (birthPlace, United States),
and (genre, Hip hop music).
5.4 Error Analysis
With Tab. 4, we provide an analysis of error type frequencies for
the rule-based approach on the basis of the evaluated sample. (1) is
caused by the entity linker, mostly due to incorrect entity borders.
For example, the tagger identies only a part of an album title. (2)
is caused by errors of the subject entity identication approach,
e.g. when the approach identies the wrong column of a table as
the one that holds subject entities. (3) can have multiple reasons,
but most often the applied rule is over-generalized (e.g. implying
when the listing is actually about athletes in
general) or applied to the wrong listing (i.e., the context described
by the rule is not expressive enough). Finally, (4) happens, for
example, when a table holds the specications of a camera as this
cannot be expressed with the given set of predicates.
Overall, most of the errors are produced by incorrectly applied
rules. This is, however, unavoidable to a certain extent as knowledge
graphs are not error-free and the data corpus is not perfect. A
substantial portion of errors is also caused by incorrectly parsed
or identied subject entities. Reducing these errors can also have a
positive impact on the generated rules as correct information about
entities is a requirement for correct rules.
In this work, we demonstrate the potential of exploiting co-occurring
similar entities for information extraction, and especially for the
discovery of novel entities. We show that it is possible to mine
expressive descriptive rules for listings in Wikipedia which can be
used to extract information about millions of novel entities.
Table 4: Error types partitioned by cause. The occurrence
values are given as their relative frequency (per 100) in the
samples evaluated in Tab. 3.
Error type Type Relation
(1) Entity parsed incorrectly 2.6 0.2
(2) Wrong subject entity identied 1.4 1.6
(3) Rule applied incorrectly 3.7 2.3
(4) Semantics of listing too complex 0.3 0.0
To improve our approach, we are investigating more sophistic-
ated ltering approaches for the generated assertions to reduce the
margin from raw to ltered assertions (see Tab. 2). Furthermore, we
are experimenting with more expressive rules (e.g. by including ad-
ditional context like substring patterns or section text) to improve
our Wikipedia-based approach.
At the moment, we extract entities from single pages. While
entity disambiguation on single pages is quite simple (on a single
Wikipedia page, it is unlikely that the same surface form refers
to dierent entities), the disambiguation entities across pages is
a much more challenging problem. Here, entity matching across
pages is required, which should, ideally, combine signals from the
source pages as well as constraints from the underlying ontology.
Furthermore, we work towards applying our approach to addi-
tional data corpora. Since the only language-dependent ingredient
of our approach is the named entity tagging, and the entity tagger
we use in our experiments has models for various languages,
our approach can also be extended to various language editions of
Besides Wikipedia, we want to apply the approach to wikis in
the Fandom
universe containing more than 380k wikis on various
domains (among them many interesting wikis for our approach, like
for example WikiLists
). For background knowledge, we plan to
rely on existing knowledge graphs in this domain like DBkWik [
or TiFi [
]. In the longer term, we want to extend the applicability
of the approach towards arbitrary web pages, using microdata and
RDFa annotations [20] as hooks for background knowledge.
Rakesh Agrawal, Tomasz Imieliński, and Arun Swami. 1993. Mining association
rules between sets of items in large databases. In 1993 ACM SIGMOD international
conference on Management of data. 207–216.
Matteo Cannaviccio, Lorenzo Ariemma, Denilson Barbosa, and Paolo Merialdo.
2018. Leveraging Wikipedia table schemas for knowledge graph augmentation.
In 21st International Workshop on the Web and Databases. 1–6.
Cuong Xuan Chu, Simon Razniewski, and Gerhard Weikum.2019. TiFi: Taxonomy
Induction for Fictional Domains. In The World Wide Web Conference. 2673–2679.
Luciano Del Corro and Rainer Gemulla. 2013. Clausie: clause-based open inform-
ation extraction. In The World Wide Web Conference. 355–366.
Xin Luna Dong, Evgeniy Gabrilovich, Kevin Murphy, Van Dang, Wilko Horn,
Camillo Lugaresi, Shaohua Sun, and Wei Zhang. 2015. Knowledge-Based Trust:
Estimating the Trustworthiness of Web Sources. VLDB Endowment 8, 9 (2015),
Michael Färber, Achim Rettinger, and Boulos El Asmar. 2016. On emerging entity
detection. In European Knowledge Acquisition Workshop. Springer, 223–238.
Information Extraction From Co-Occurring Similar Entities WWW ’21, April 19–23, 2021, Ljubljana, Slovenia
Besnik Fetahu, Avishek Anand, and Maria Koutraki. 2019. TableNet: An approach
for determining ne-grained relations for Wikipedia tables. In The World Wide
Web Conference. 2736–2742.
Luis Galárraga, Christina Teioudi, Katja Hose, and Fabian M Suchanek. 2015.
Fast rule mining in ontological knowledge bases with AMIE+. The VLDB Journal
24, 6 (2015), 707–730.
Nicolas Heist, Sven Hertling, Daniel Ringler, and Heiko Paulheim. 2020. Know-
ledge Graphs on the Web–an Overview. Studies on the Semantic Web 47 (2020),
Nicolas Heist and Heiko Paulheim. 2019. Uncovering the Semantics of Wikipedia
Categories. In International Semantic Web Conference. Springer, 219–236.
Nicolas Heist and Heiko Paulheim. 2020. Entity Extraction from Wikipedia List
Pages. In Extended Semantic Web Conference. Springer, 327–342.
Sven Hertling and Heiko Paulheim. 2020. DBkWik: extracting and integrating
knowledge from thousands of wikis. Knowledge and Information Systems 62, 6
(2020), 2169–2190.
Jens Lehmann. 2009. DL-Learner: learning concepts in description logics. The
Journal of Machine Learning Research 10 (2009), 2639–2642.
Jens Lehmann et al
2015. DBpedia–a large-scale, multilingual knowledge base
extracted from Wikipedia. Semantic Web 6, 2 (2015), 167–195.
Oliver Lehmberg and Christian Bizer. 2017. Stitching web tables for improving
matching quality. VLDB Endowment 10, 11 (2017), 1502–1513.
Guiliang Liu, Xu Li, Jiakang Wang, Mingming Sun, and Ping Li. 2020. Extracting
Knowledge from Web Text with Monte Carlo Tree Search. In The Web Conference
2020. 2585–2591.
Erin Macdonald and Denilson Barbosa. 2020. Neural Relation Extraction on
Wikipedia Tables for Augmenting Knowledge Graphs. In 29th ACM International
Conference on Information & Knowledge Management. 2133–2136.
Christian Meilicke, Melisachew Wudage Chekol, Daniel Runelli, and Heiner
Stuckenschmidt. 2019. Anytime Bottom-Up Rule Learning for Knowledge Graph
Completion.. In 28th International Joint Conference on Articial Intelligence. 3137–
Pablo N Mendes et al
2011. DBpedia spotlight: shedding light on the web of
documents. In 7th international conference on semantic systems. 1–8.
Robert Meusel, Petar Petrovski, and Christian Bizer. 2014. The webdatacommons
microdata, rdfa and microformat dataset series. In International Semantic Web
Conference. Springer, 277–292.
Yaser Oulabi and Christian Bizer. 2019. Using weak supervision to identify
long-tail entities for knowledge base completion. In International Conference on
Semantic Systems. Springer, 83–98.
Heiko Paulheim. 2017. Knowledge graph renement: A survey of approaches
and evaluation methods. Semantic web 8, 3 (2017), 489–508.
Heiko Paulheim and Johannes Fümkranz. 2012. Unsupervised generation of data
mining features from linked open data. In 2nd international conference on web
intelligence, mining and semantics. 1–12.
Heiko Paulheim and Simone Paolo Ponzetto. 2013. Extending DBpedia with
Wikipedia List Pages. In 1st International Workshop on NLP and DBpedia, Vol. 1064.
CEUR Workshop Proceedings, 85–90.
Dominique Ritze, Oliver Lehmberg, Yaser Oulabi, and Christian Bizer. 2016.
Proling the potential of web tables for augmenting cross-domain knowledge
bases. In The World Wide Web Conference. 251–261.
Ahmad Sakor, Kuldeep Singh, Anery Patel, and Maria-Esther Vidal. 2020. Falcon
2.0: An entity and relation linking tool over Wikidata. In 29th ACM International
Conference on Information & Knowledge Management. 3141–3148.
Gabriel Stanovsky, Julian Michael, Luke Zettlemoyer, and Ido Dagan. 2018. Su-
pervised open information extraction. In 2018 Conference of the North American
Chapter of the Association for Computational Linguistics: Human Language Tech-
nologies, Volume 1 (Long Papers). 885–895.
Fabian M Suchanek, Gjergji Kasneci, and Gerhard Weikum. 2007. Yago: a core of
semantic knowledge. In The World Wide Web Conference. 697–706.
Marieke Van Erp, Pablo Mendes, Heiko Paulheim, Filip Ilievski, Julien Plu, Gi-
useppe Rizzo, and Jörg Waitelonis. 2016. Evaluating entity linking: An analysis
of current benchmark datasets and a roadmap for doing a better job. In Proceed-
ings of the Tenth International Conference on Language Resources and Evaluation
(LREC’16). 4373–4379.
Johanna Völker and Mathias Niepert. 2011. Statistical schema induction. In
Extended Semantic Web Conference. Springer, 124–138.
Denny Vrandečić and Markus Krötzsch. 2014. Wikidata: a free collaborative
knowledgebase. Commun. ACM 57, 10 (2014), 78–85.
Quan Wang, Zhendong Mao, Bin Wang, and Li Guo. 2017. Knowledge graph
embedding: A survey of approaches and applications. IEEE Transactions on
Knowledge and Data Engineering 29, 12 (2017), 2724–2743.
Bo Xu, Chenhao Xie, Yi Zhang, Yanghua Xiao, Haixun Wang, and Wei Wang. 2016.
Learning dening features for categories. In 25th International Joint Conference
on Articial Intelligence. 3924–3930.
Mohamed Yakout, Kris Ganjam, Kaushik Chakrabarti, and Surajit Chaudhuri.
2012. InfoGather: entity augmentation and attribute discovery by holistic match-
ing with web tables. In 2012 ACM SIGMOD International Conference on Manage-
ment of Data. 97–108.
Shuo Zhang et al
2020. Novel Entity Discovery from Web Tables. In The Web
Conference 2020. 1298–1308.
Shuo Zhang and Krisztian Balog. 2020. Web Table Extraction, Retrieval, and
Augmentation: A Survey. ACM Transactions on Intelligent Systems and Technology
(TIST) 11, 2 (2020), 1–35.
Shuo Zhang, Krisztian Balog, and Jamie Callan. 2020. Generating Categories for
Sets of Entities. In 29th ACM International Conference on Information & Knowledge
Management. 1833–1842.
Ziqi Zhang. 2014. Towards ecient and eective semantic table interpretation.
In International Semantic Web Conference. Springer, 487–502.
Ganggao Zhu and Carlos A Iglesias. 2018. Exploiting semantic similarity for
named entity disambiguation in knowledge graphs. Expert Systems with Applica-
tions 101 (2018), 8–24.
ResearchGate has not been able to resolve any citations for this publication.
Conference Paper
Full-text available
The Natural Language Processing (NLP) community has significantly contributed to the solutions for entity and relation recognition from a natural language text, and possibly linking them to proper matches in Knowledge Graphs (KGs). Considering Wikidata as the background KG, there are still limited tools to link knowledge within the text to Wikidata. In this paper, we present Falcon 2.0, the first joint entity and relation linking tool over Wikidata. It receives a short natural language text in the English language and outputs a ranked list of entities and relations annotated with the proper candidates in Wikidata. The candidates are represented by their Internationalized Resource Identifier (IRI) in Wikidata. Falcon 2.0 resorts to the English language model for the recognition task (e.g., N-Gram tiling and N-Gram splitting), and then an optimization approach for the linking task. We have empirically studied the performance of Falcon 2.0 on Wikidata and concluded that it outperforms all the existing baselines. Falcon 2.0 is open source and can be reused by the community; all the required instructions of Falcon 2.0 are well-documented at our GitHub repository ( We also demonstrate an online API, which can be run without any technical expertise. Falcon 2.0 and its background knowledge bases are available as resources at
Full-text available
Tables are powerful and popular tools for organizing and manipulating data. A vast number of tables can be found on the Web, which represent a valuable knowledge resource. The objective of this survey is to synthesize and present two decades of research on web tables. In particular, we organize existing literature into six main categories of information access tasks: table extraction, table interpretation, table search, question answering, knowledge base augmentation, and table augmentation. For each of these tasks, we identify and describe seminal approaches, present relevant resources, and point out interdependencies among the different tasks.
Full-text available
Popular cross-domain knowledge graphs, such as DBpedia and YAGO, are built from Wikipedia, and therefore similar in coverage. In contrast, Wikifarms like Fandom contain Wikis for specific topics, which are often complementary to the information contained in Wikipedia, and thus DBpedia and YAGO. Extracting these Wikis with the DBpedia extraction framework is possible, but results in many isolated knowledge graphs. In this paper, we show how to create one consolidated knowledge graph, called DBkWik, from thousands of Wikis. We perform entity resolution and schema matching, and show that the resulting large-scale knowledge graph is complementary to DBpedia. Furthermore, we discuss the potential use of DBkWik as a benchmark for knowledge graph matching.
Conference Paper
We propose an anytime bottom-up technique for learning logical rules from large knowledge graphs. We apply the learned rules to predict candidates in the context of knowledge graph completion. Our approach outperforms other rule-based approaches and it is competitive with current state of the art, which is based on latent representations. Besides, our approach is significantly faster, requires less computational resources, and yields an explanation in terms of the rules that propose a candidate.
Conference Paper
Taxonomies are important building blocks of structured knowledge bases, and their construction from text sources and Wikipedia has received much attention. In this paper we focus on the construction of taxonomies for fictional domains, using noisy category systems from fan wikis or text extraction as input. Such fictional domains are archetypes of entity universes that are poorly covered by Wikipedia, such as also enterprise-specific knowledge bases or highly specialized verticals. Our fiction-targeted approach, called TiFi, consists of three phases: (i) category cleaning, by identifying candidate categories that truly represent classes in the domain of interest, (ii) edge cleaning, by selecting subcategory relationships that correspond to class subsumption, and (iii) top-level construction, by mapping classes onto a subset of high-level WordNet categories. A comprehensive evaluation shows that TiFi is able to construct taxonomies for a diverse range of fictional domains such as Lord of the Rings, The Simpsons or Greek Mythology with very high precision and that it outperforms state-of-the-art baselines for taxonomy induction by a substantial margin.
Conference Paper
We focus on the problem of interlinking Wikipedia tables with fine-grained table relations: equivalent and subPartOf. Such relations allow us to harness semantically related information by accessing related tables or facts therein. Determining the type of a relation is not trivial. Relations are dependent on the schemas, the cell-values, and the semantic overlap of the cell values in tables. We propose TableNet, an approach for interlinking tables with subPartOf and equivalent relations. TableNet consists of two main steps: (i) for any source table we provide an efficient algorithm to find candidate related tables with high coverage, and (ii) a neural based approach that based on the table schemas and data, determines with high accuracy the fine-grained relation. Based on an extensive evaluation with more than 3.2M tables, we show that TableNet retains more than 88% of relevant tables pairs, and assigns table relations with an accuracy of 90%.