Entity linking with a knowledge base issues,

ENTITY LINKING WITH A KNOWLEDGE BASE: ISSUES,
TECHNIQUES, AND SOLUTIONS
Abstract—The large number of potential applications from bridging web data
with knowledge bases have led to an increase in the entity linking research. Entity
linking is the task to link entity mentions in text with their corresponding entities in
a knowledge base. Potential applications include information extraction,
information retrieval, and knowledge base population. However, this task is
challenging due to name variations and entity ambiguity. In this survey, we present
a thorough overview and analysis of the main approaches to entity linking, and
discuss various applications, the evaluation of entity linking systems, and future
directions.
EXISTING SYSTEM:
The entity linking task is challenging due to name variations and entity ambiguity.
A named entity may have multiple surface forms, such as its full name, partial
names, aliases, abbreviations, and alternate spellings. For example, the named
entity of “Cornell University” has its abbreviation “Cornell” and the named entity
of “New York City” has its nickname “Big Apple”. An entity linking system has to
identify the correct mapping entities for entity mentions of various surface forms.

On the other hand, an entity mention could possibly denote different named
entities. For instance, the entity mention “Sun” can refer to the star at the center of
the Solar System, a multinational computer company, a fictional character named
“Sun-Hwa Kwon” on the ABC television series “Lost” or many other entities
which can be referred to as “Sun”. An entity linking system has to disambiguate
the entity mention in the textual context and identify the mapping entity for each
entity mention.
PROPOSED SYSTEM:
We have presented a comprehensive survey for entity linking. Specifically, we
have surveyed the main approaches utilized in the three modules of entity linking
systems (i.e., Candidate Entity Generation, Candidate Entity Ranking, and
Unlinkable Mention Prediction), and also introduced other critical aspects of entity
linking such as applications, features, and evaluation. Although there are so many
methods proposed to deal with entity linking, it is currently unclear which
techniques and systems are the current state-of-the-art, as these systems all differ
along multiple dimensions and are evaluated over different data sets. A single
entity linking system typically performs very differently for different data sets and
domains. Although the supervised ranking methods seem to perform much better
than the unsupervised approaches with respect to candidate entity ranking, the
overall performance of the entity linking system is also significantly influenced by

techniques adopted in the other two modules (i.e., Candidate Entity Generation and
Unlinkable Mention Prediction). Supervised techniques require many annotated
training examples and the task of annotating examples is costly. Furthermore, the
entity linking task is highly data dependent and it is unlikely a technique dominates
all others across all data sets. For a given entity linking task, it is difficult to
determine which techniques are best suited.
Module 1
Entity lining System
Entity linking system consists of the following three modules:
_ Candidate entity generation. In this module, for each entity mention m 2 M, the
entity linking system aims to filter out irrelevant entities in the knowledge base and
retrieve a candidate entity set Em which contains possible entities that entity
mention m may refer to. To achieve this goal, a variety of techniques have been
utilized by some state-of-the-art entity linking systems, such as name dictionary
based techniques, surface form expansion from the local document, and methods
based on search engine.
_ Candidate entity ranking. In most cases, the size of the candidate entity set Em is
larger than one. Researchers leverage different kinds of evidence to rank the
candidate entities in Em and try to find the entity e 2 Em which is the most likely
link for mention m. To deal with the problem of predicting unlinkable mentions,

some work leverages this module to validate whether the topranked entity
identified in the Candidate Entity Ranking module is the target entity for mention
m. Otherwise, they return NIL for mention m. An overview of the main
approaches for predicting unlinkable mentions.
Module 2
Candidate entity generation
Entity Generation module, for each entity mention m 2 M, entity linking systems
try to include possible entities that entity mention m may refer to in the set of
candidate entities Em. Approaches to candidate entity generation are mainly based
on string comparison between the surface form of the entity mention and the name
of the entity existing in a knowledge base. This module is as important as the
Candidate Entity Ranking module and critical for a successful entity linking
system according to the experiments conducted by Hachey et al. [33]. In the
remainder of this section, we review the main approaches that have been applied
for generating the candidate entity set Em for entity mention m.
Module 3
Name Dictionary BasedTechniques

Name dictionary based techniques are the main approaches to candidate entity
generation and are leveraged by many entity linking systems.The structure of
Wikipedia provides a set of useful features for generating candidate entities, such
as entity pages, redirect pages, disambiguation pages, bold phrases from the first
paragraphs, and hyperlinks in Wikipedia articles. These entity linking systems
leverage different combinations of these features to build an offline name
dictionary D between various names and their possible mapping entities, and
exploit this constructed name dictionary D to generate candidate entities. This
name dictionary D contains vast amount of information about various names of
named entities, like name variations, abbreviations, confusable names, spelling
variations, nicknames, etc. Specifically, the name dictionary D is a hkey, valuei
mapping, where the key column is a list of names. Suppose k is a name in the key
column, and its mapping value k:value in the value column is a set of named
entities which could be referred to as the name k. The dictionary D is constructed
by leveraging features from Wikipedia as follows:
_ Entity pages. Each entity page in Wikipedia describes a single entity and
contains the information focusing on this entity. Generally, the title of each page is
the most common name for the entity described in this page, e.g., the page title
“Microsoft” for that giant software company headquartered in Redmond. Thus, the

title of the entity page is added to the key column inD as a name k, and the entity
described in this page is addedas k:value.
_ Redirect pages. A redirect page exists for each alternative name which could be
used to refer to an existing entity in Wikipedia. For example, the article titled
“Microsoft Corporation” which is the full name of Microsoft contains a pointer to
the article of the entity Microsoft. Redirect pages often indicate synonym terms,
abbreviations, or other variations of the pointed entities. Therefore, the title of the
redirect page is added to the key column in D as a name , and the pointed entity is
added as k:value.
_ Disambiguation pages. When multiple entities in Wikipedia could be given the
same name, a disambiguation page is created to separate them and contains a list of
references to those entities. For example, the disambiguation page for the name
“Michael Jordan” lists thirteen associated entities having the same name of
“Michael Jordan” including the famous NBA player and the Berkeley professor.
These disambiguation pages are very useful in extracting abbreviations or other
aliases of entities. For each disambiguation page, the title of this page is added to
the key column in D as a name k, and the entities listed in this page are added as
k:value.
_ Bold phrases from the first paragraphs. In general, the first paragraph of a
Wikipedia article is a summary of the whole article. It sometimes contains a few

phrases written in bold. Varma et al. observed that these bold phrases invariably
are nick names, alias names or full names of the entity described in this paper. For
instance, in the first paragraph of the entity page of Hewlett-Packard (HP), there
are two phrases written in bold (i.e., “Hewlett-Packard Company” and “HP”)
which are respectively the full name and the abbreviation for the entity Hewlett-
Packard. Thus, for each of the bold phrases in the first paragraph of each
Wikipedia page, it is added to the key column in D as a name k, and the entity
described in this page is added as k:value.
_ Hyperlinks in Wikipedia articles. An article in Wikipedia often contains
hyperlinks which link to the pages of the entities mentioned in this article. The
anchor text of a link pointing to an entity page provides a very useful source of
synonyms and other name variations of the pointed entity, and could be regarded
as a name of that linked entity. For example, in the entity page of Hewlett-Packard,
there is a hyperlink pointing to the entity William Reddington Hewlett whose
anchor text is “Bill Hewlett”, which is an alias name of the entity William
Reddington Hewlett. Hence, the anchor text of the hyperlink is added to the key
column in D as a name k, and the pointed entity is added as k:value. Using these
features from Wikipedia described above, entity linking systems could construct a
dictionary D. Besides leveraging the features from Wikipedia, there are some

studies that exploit query click logs and web documents to find entity synonyms,
which are also helpful for the name dictionary construction.
Module 4
Surface Form Expansion from the Local Document
Since some entity mentions are acronyms or part of their full names, one category
of entity linking systems use the surface form expansion techniques to identify
other possible expanded variations (such as the full name) from the associated
document where the entity mention appears. Then they could leverage these
expanded forms to generate the candidate entity set using other methods such as
the name dictionary based techniques introduced above. We categorize the surface
form expansion techniques into the heuristic based methods and the supervised
learning methods.
Module 5
Candidate entity ranking
In the previous section, we described methods that could generate the candidate
entity set Em for each entity mention m. We denote the size of Em as jEmj, and
use 1 _ i _ jEmj to index the candidate entity in Em. The candidate entity with
index i in Em is denoted by ei. In most cases, the size of the candidate entity set

Em is larger than one. For instance, Ji et al. [89] showed that the average number
of candidate entities per entity mention on the TAC-KBP2010 data set is 12.9, and
this average number on the TAC-KBP2011 data set is 13.1. In addition, this
average number is 73 on the CoNLL data set utilized in [58]. Therefore, the
remaining problem is how to incorporate different kinds of evidence to rank the
candidate entities in Em and pick the proper entity from Em as the mapping entity
for the entity mention m. The Candidate Entity Ranking module is a key
component for the entity linking system. We can broadly divide these candidate
entity ranking methods into two categories:
_ Supervised ranking methods. These approaches rely on annotated training data
to “learn” how to rank the candidate entities in Em. These approaches include
binary classification methods, learning to rank methods, probabilistic methods, and
graph based approaches.
_ Unsupervised ranking methods. These approaches are based on unlabeled corpus
and do not require any manually annotated corpus to train the model. These
approaches include vector space model (VSM) based methods and information
retrieval based methods. In this section, all candidate entity ranking methods are
illustrated according to the above categorization. In addition, we could also
categorize the candidate entity ranking methods into another three categories:

_ Independent ranking methods. These approaches consider that entity mentions
which need to be linked in a document are independent, and do not leverage the
relations between the entity mentions in one document to help candidate entity
ranking. In order to rank the candidate entities, they mainly leverage the context
similarity between the text around the entity mention and the document associated
with the candidate entity.
_ Collective ranking methods. These methods assume that a document largely
refers to coherent entities from one or a few related topics, and entity assignments
for entity mentions in one document are interdependent with each other. Thus, in
these methods, entity mentions in one document are collectively linked by
exploiting this “topical coherence”.
_ Collaborative ranking methods. For an entity mention that needs to be linked,
these approaches identify other entity mentions having similar surface forms and
similar textual contexts in the other documents. They leverage this cross-document
extended context information obtained from the other similar entity mentions and
the context information of the entity mention itself to rank candidate entities for
the entity mention.
CONCLUSION:

In this paper, we have presented a comprehensive survey for entity linking.
Specifically, we have surveyed the main approaches utilized in the three modules
of entity linking systems (i.e., Candidate Entity Generation, Candidate Entity
Ranking, and Unlinkable Mention Prediction), and also introduced other critical
aspects of entity linking such as applications, features, and evaluation. Although
there are so many methods proposed to deal with entity linking, it is currently
unclear which techniques and systems are the current state-of-the-art, as these
systems all differ along multiple dimensions and are evaluated over different data
sets. A single entity linking system typically performs very differently for different
data sets and domains. Although the supervised ranking methods seem to perform
much better than the unsupervised approaches with respect to candidate entity
ranking, the overall performance of the entity linking system is also significantly
influenced by techniques adopted in the other two modules (i.e., Candidate Entity
Generation and Unlinkable Mention Prediction). Supervised techniques require
many annotated training examples and the task of annotating examples is costly.
Furthermore, the entity linking task is highly data dependent and it is unlikely a
technique dominates all others across all data sets. For a given entity linking task,
it is difficult to determine which techniques are best suited. There are many aspects
that affect the design of the entity linking system, such as the system requirement
and the characteristics of the data sets. Although our survey has presented many

efforts in entity linking, we believe that there are still many opportunities for
substantial improvement in this field. In the following, we point out some
promising research directions in entity linking.
REFERENCES
[1] S. Auer, C. Bizer, G. Kobilarov, J. Lehmann, and Z. Ives, “DBpedia: A nucleus
for a web of open data,” in Proc. 6th Int. Semantic Web 2nd Asian Conf. Asian
Semantic Web Conf., 2007, pp. 11–15.
[2] F. M. Suchanek, G. Kasneci, and G. Weikum, “Yago: A core of semantic
knowledge unifying wordnet and wikipedia,” in Proc. 16th Int. Conf. World Wide
Web, 2007, pp. 697–706.
[3] K. Bollacker, C. Evans, P. Paritosh, T. Sturge, and J. Taylor, “Freebase: A
collaboratively created graph database for structuring human knowledge,” in Proc.
ACM SIGMOD Int. Conf. Manage. Data, 2008, pp. 1247–1250.
[4] O. Etzioni, M. Cafarella, D. Downey, S. Kok, A.-M. Popescu, T. Shaked, S.
Soderland, D. S. Weld, and A. ates, “Web-scale information extraction in
knowitall: (preliminary results),” in Proc. 13th Int. Conf. World Wide Web, 2004,
pp. 100–110.
[5] A. Carlson, J. Betteridge, R. C. Wang, E. R. Hruschka, Jr, and T. M. Mitchell,
“Coupled semi-supervised learning for information extraction,” in Proc. 3rd ACM
Int. Conf. Web Search Data Mining, 2010, pp. 101–110.

[6] W. Wu, H. Li, H. Wang, and K. Q. Zhu, “Probase: A probabilistic taxonomy
for text understanding,” in Proc. ACM SIGMOD Int. Conf. Manage. Data, 2012,
pp. 481–492.
[7] T. Berners-Lee, J. Hendler, and O. Lassila, “The semantic web,” Sci. Am., vol.
284, pp. 34–43, 2001.
[8] E. Agichtein and L. Gravano, “Snowball: Extracting relations from large plain-
text collections,” in Proc. ACM Int. Conf. Digital Libraries, 2000, pp. 85–94.
[9] D. Zelenko, C. Aone, and A. Richardella, “Kernel methods for relation
extraction,” J. Mach. Learn. Res., vol. 3, pp. 1083–1106, 2003.
[10] T. Hasegawa, S. Sekine, and R. Grishman, “Discovering relations among
named entities from large corpora,” in Proc. 42nd Ann. Meeting Assoc. Comput.
Linguistics, 2004, pp. 415–422.

Entity linking with a knowledge base issues,

More Related Content

What's hot (20)

Viewers also liked (16)

Similar to Entity linking with a knowledge base issues, (20)

More from Nexgen Technology (20)

Entity linking with a knowledge base issues,