Named Entity Recognition Using Web Document Corpus

International Journal of Managing Information Technology (IJMIT) Vol.3, No.1, February 2011
DOI : 10.5121/ijmit.2011.3104 46
NAMED ENTITY RECOGNITION USING
WEB
DOCUMENT CORPUS
WAHIBA BEN ABDESSALEM KARAA
Institut Superieur de Gestion de Tunis 41 avenue de la Liberté
Cité Bouchoucha, Le Bardo 2000, Tunisia,
wahiba.abdessalem@isg.rnu.tn
Abstract
This paper introduces a named entity recognition approach in textual corpus. This Named Entity (NE)
can be a named: location, person, organization, date, time, etc., characterized by instances. A NE is
found in texts accompanied by contexts: words that are left or right of the NE. The work mainly aims at
identifying contexts inducing the NE’s nature. As such, The occurrence of the word "President" in a
text, means that this word or context may be followed by the name of a president as President
"Obama". Likewise, a word preceded by the string "footballer" induces that this is the name of a
footballer. NE recognition may be viewed as a classification method, where every word is assigned to
a NE class, regarding the context.
The aim of this study is then to identify and classify the contexts that are most relevant to recognize a
NE, those which are frequently found with the NE. A learning approach using training corpus: web
documents, constructed from learning examples is then suggested. Frequency representations and
modified tf-idf representations are used to calculate the context weights associated to context
frequency, learning example frequency, and document frequency in the corpus.
Keywords: Named entity; Learning; Information extraction; tf-idf; Web document.
1.Introduction
Named Entity Recognition is complex in various areas of automatic Natural Language
Processing of (NLP), document indexing, document annotation, translation, etc. [2]. It is a
fundamental step in various Information Extraction (IE) tasks. It has been an essential task in
several research teams such as the Message Understanding Conferences (MUC), the Conferences
on Natural Language Learning (CoNLL), etc.
Named entities (NE) are phrases containing the names of persons, locations, etc. They are
particularly important for the access to document content, since they form the building blocks
upon which the analysis of documents is based.
This paper discusses the use of learning approach for the problem of NE recognition. The
goal is to reveal contextual NE in a document corpus. A context considers words surrounding the
NE in the sentence in which it appears, it is a sequence of words, that are left or right of the NE.
We use, in this work, the makings of learning technologies, combined with statistical models
[17], to extract contexts from Web document corpus, to identify the most pertinent contexts for
the recognition of a NE. We investigate the impact of using different feature weighting measures,
in the hope that they will yield more context classification.

47
The remaining of the paper is organized as follows: section 2, introduces the state of the art of
methods applied in Named Entity Recognition. Section 3 describes the methodology and section
4 gives test results of our approach. Section 5 is devoted the work’s conclusion.
2.State of the art
Named entity recognition can be used to perform numerous processing tasks in various areas:
Information extraction systems [6], text mining [8], [16], Automatic Speech Recognition (ASR)
[5], etc.
Several works are particularly interested in the recognition of named entities. Mikheev et al.
[12] have built a system for recognizing named entities, which combines a model based on
grammar rules, and statistical models, without resorting to named entity lists.
Collins et al. [1] suggests an algorithm for named entity classification, based on the meaning
word disambiguation, and exploits the redundancy in the contextual characteristics.
This system operates a large corpus to produce a generic list of proper nouns. The names are
collected by searching for a syntax diagram with specific properties. For example, a proper name
is a sequence of consecutive words in a nominal phrase, etc.
Petasis et al. [14] presented a method that helps to build a rules-based system for recognition
and classification of named entities. They have used machine learning, to monitor system
performance and avoid manual marking.
In his paper [10], Mann explores the idea of fine-grained proper noun ontology and its use in
question answering. The ontology is built from unrestricted text using simple textual co-
occurrence patterns. This ontology is therefore used on a question answering task to provide
primary results on the utility of this information. However, this method has a low coverage.
The Nemesis system presented by Fourour [6] is founded on some heuristics, allowing the
identification of named entities, and their classification by detecting the boundaries of the entity
called "context" to the left or right, and by studying syntactic, or morphological of these entities.
For example, acronyms are named entities consisting of a single lexical unit comprising several
capital letters, etc.
Krstev et al [9] suggested a basic structure of a relational model of a multilingual dictionary of
proper names based on four-level ontology. However, the implementation is not yet completed, it
is only expected.
The KNOWITALL system planned by Etzioni et al [4] aims at automating the process of
extracting named entities from the Web in an unsupervised and scalable manner. This system is
not intended for recognizing a named entity, but used to create long lists of named entities.
However, it is not designed to resolve the ambiguity in some documents.
Friburger [7] recommends a method based on rules for finding a large proportion of person
names. However, this method has some limitations as errors, and missing responses.

48
Nadeau et al. [13] have suggested a system for recognizing named entities. Their work is
based on those of Collins et al. [1], and Etzioni et al. [4]. The system exploits human-generated
HTML markup in Web pages to generate gazetteers, then it uses simple heuristics for the entity
disambiguation in the context of a given document.
Marthineau et al. [11] showed the validity of existing local grammars in the system
GRAALWEB to recognize and extract the named entities.
In view of works touching the recognition of named entities, we perceive that most of them
are based on a set of rules in relation to predefined categories: morphological, grammatical, etc.
[7], [11], [14], or on predefined lists of dictionaries [9]. The ontology domain is still in
exploration [10].
We adopted the idea of Nemesis based on the left and the right context of the named entity.
However, our approach does not mark the context derived from syntactic or morphological rules,
but identifies the context founded on learning phase. The objective is thus to carry out a system,
able to induce the nature of a named entity, following the meeting of certain indicators, and this,
in any language, without requiring to dictionaries or lists of named entities.
For our approach we use, for the learning phase, documents resulting from the Web. We use
context frequency representations, based on tf-idf version, to calculate context weights. These
weights are used to determine the most pertinent contexts, to identify the named entity nature.
3.Methodology
The system encloses three main phases. The first is the training corpus collection. The second
is the context extraction and the contexts classification, according to weighting measures. The
third phase is the weighting measures exploitation, in order to build a model for NE recognition.
The figure (Fig 1) summarizes the architecture.

49
Fig. 1. The architeture
3.1Training corpus construction
The first step is to build an initial corpus containing web documents. This corpus is called
learning corpus. Our algorithm proceeds as follows:
(i) We Furnish a set of learning examples, which are instances of a named entity class. For
example for the class entity "disease", we provide as learning examples, names of diseases
such as "flu", "Eczema", "measles", etc. For the named entity class "football player" must be
provided as instances, footballer names such as "platini", "Zidane", "Maradona", etc.
(ii) For all the instances a query is sent to a Web search engine (Yahoo, Google, AltaVista,
etc.), by using a specific API related to each engine, searching for context string
surrounding the instances: words before, or/and after the instance, the length of the string is
an input parameter.
(iii) The result is a list of links. Afterwards, for each link, a new request is released in order to
retrieve and save the document. To increase the number of returned links, the user should
provide a maximum of learning examples. The figure (Fig 2) is an extract of the API used
with yahoo engine:

50
Fig. 2. The API algorithm
3.2 Learning phase
A context is a set of words preceding, or following a named entity. We focus, in this study, on the
words that precede a named entity instance; we consider that the left context is more relevant then
the right one.
For instance, we suggest the concept "President" with the six instances, considered as learning
examples: "Bush", "G W. Bush", "Nicolas Sarkozy," Sarkozy "," Chirac ", and "Jacques Chirac".
The result of the extraction phase is a set of context candidates. The figure (Fig. 3.) is an extract
of a document where the context candidates are pointed:
Fig. 3.Example of extracted document
import com.yahoo.search.SearchClient;
import com.yahoo.search.SearchException;
import com.yahoo.search.WebSearchRequest;
import com.yahoo.search.WebSearchResult;
import com.yahoo.search.WebSearchResults;
public class Extract_Req_Thread extends Thread {
. . .
SearchClient client = new SearchClient("Yahoo key");
WebSearchRequest request = new WebSearchRequest(“instance of the named entity”);
WebSearchResults results = client.webSearch(request);
. . .
for (int i = 0; i < results.listResults().length;i++) {
WebSearchResult result = results.listResults()[i];
//Traitements
}
…At age 69, Jacques Chirac faced fourth…a leading voice against U.S. President George
W. Bush's administration's conduct towards Iraq. Despite … supporting the decisions of
President Bush and Prime Minister Tony Blair. See also …the second round of elections,
instead, Chirac faced controversial far right politician … of the unity of our nation, Chirac
said before the presidential…during Bastille Day celebration, Chirac survived an
association attempt … at the invitation of president Nicolas Sarkosy with whom he met on
Wednesday … Vice president Dick Cheney looks on while U.S. President Dick Cheney
looks on while U.S. President Bush during an alternative …

51
There are eight instances with different contexts. We can notice that the context "President" is
most common with the learning examples, which makes it the more appropriate to recognize the
name of a president. However, the context occurrence number is an insufficient parameter to
decide that it is the best context to construct a rule:
President <president-name>.
The goal is to find high-quality context. To define the context's quality, we observe some
metrics calculated through frequencies, given that instances found on the Web surrounded by the
context.
Estimating pertinence of contexts is not sufficient by only the Frequency of these contexts
found surrounding a NE instance, especially when we have no labelled negative learning
examples, but only positive ones.
The purpose of this work is to identify the most pertinent contexts for the identification of a
given named entity. For this intention we calculate weights to classify contexts taking into
account the context frequency regarding learning examples, the inverse context frequency, the
learning example frequency in the document corpus, and the frequency of documents containing
contexts in the document corpus. To calculate context weights, we use context frequency
representations based on tf-idf version. tfidf (Term Frequency-Inverse Document Frequency) is
one of the most classics and most common weighting method used to describe documents in the
Vector Space Model [17], particularly in Information Retrieval (IR). The tf considers the term
frequency in the document: the more a word occurs in a document, the more it is expected to be
significant in this document. In addition, idf inverse document frequency measures the term
frequency in the corpus: the more a word appears in a corpus, the more it is estimated irrelevant
for the document.
The Term Trequency (tf ) of a term ti for a document dj is calculated as follows:
∑
=
k kj
ij
ij
frequency
frequency
tf
(1)
Where frequency kj is the occurrence number of the term ti in dj. The denominator is the
number of occurrences of all terms in the document dj.
The idf, Inverse Document Frequency component is computed as follows:
i
i
n
N
og
idf l
=
(2)
Where N is the total number of documents in the corpus, and ni is the number of documents in
which the term ti emerges. In tf.idf weighting is:
i
ij
ij idf
tf
w ×
=
(3)

52
Many alternatives have been suggested to the basic tfidf formula, where the tf or idf part is
modified using functions related to characteristic selection. It is in our case to measure the
importance of a context ci to a NE; in our work we propose a modified formula. We define:
Definition 1: context frequency
The context frequency (cfi ) of a context ci in a document corpus is calculated as follows:
i
i
i
nc
nc
cf
∑
=
(4)
We consider the variable nci the occurrence number of a context within a document corpus
accompanied by a NE Learning Example (LE), and ∑ i
nc
is the occurrence number of all
contexts in the document corpus.
Definition 2: learning example frequency
NLE
nle
lef i
i=
(5)
We designated nlei the number of LE located with the context ci, in the corpus, and NLE the
total number of learning examples used for training.
Definition 3: document frequency
We used the variable ndi to assign the occurrence number of documents in the corpus
containing a context ci, and coming from different sources. Di is the occurrence number of
documents in the corpus containing a context ci.
The document frequency dfi is calculated as follows:
i
i
i
D
nd
df =
(6)
Definition 4: inverse context frequency
We considered the hypothesis that the context is significant for a NE, if it does not often
appears with other phrases in the corpus. icf inverse context frequency, measures the context
frequency associated to all the phrases escorted by the context: the more a context appears in a
corpus accompanied with other phrases, the more it is estimated irrelevant for the NE.
icfi is calculated using nci, the occurrence number of a context within a document corpus
accompanied by learning example, and the variable Ci, which represents the total occurrence
number of the context accompanied with other phrases in the corpus.
i
i
i
C
nc
icf =
(7)
Definition 5: context weight
The product of the obtained frequencies provides the weight wi adapted to the detection of the
context pertinence:
i
i
i
i
i icf
df
lef
cf
w ×
×
×
=
(8)

53
3.3 Recognition model
NE recognition may be viewed as a classification problem, where every word is assigned to a NE
class, regarding the context. Since decision tree algorithms are widely used for data mining [3],
classification, etc. in this third phase, we design algorithm recognition, in the form of a decision
tree, based on C4.5 algorithm [15].
In many cases, the same context can introduce different NE. For example, the context "Mr."
precedes both a footballer name and a president name: Mr. Zidane, Mr. Obama. The intention is
to decide whether the NE is a president name or a footballer name.
We define votei for each NE with each contexti in the document. The value of votei is
incremented with the weight of the context, each time a contexti is encountered:
(9)
Afterward, the algorithm decides the nature of the NE, once the value of votei reaches a
threshold, and depending on the value of votej, …, voten reached by the other contexts.
4.Experimental results
We conducted several tests with different NE. For example for the NE class "capital", we
extracted a corpus of documents from 65 URLs, obtained by queries on the search engine Yahoo.
For the 13 following instances (or learning examples): "Paris", "Tunis", "Cairo", "Athens",
"Abuja", "Berlin", "Bucharest", "Budapest", "Brasilia", "Freetown", "Dublin", "Vienna", and
"Doha". We obtained 2398 contexts composed of two different words to the left. The learning
examples are occurred 4264 times in the corpus with the contexts. We present in the following
table an extract of the obtained context classification results:
Table 1. Context classification.
Context cf df lef icf W
Hotels in 0,0039869 0,4444445 0,5384616 8,5 0,008110106
Map of 0,0028143 0,6 0,5384616 6 0,005455413
travel to 0,000469 1 0,1538462 2 0,000144308
Guide to 0,000469 1 0,1538462 2 0,000144308
Places in 0,0011726 0,75 0,3076923 1 0,0002706
hotels in 0,0021107 0,5714286 0,6153847 3 0,002226673
map of 0,0007036 1 0,2307692 1 0,000162369
Hospital in 0,0007036 1 0,0769231 0,4 2,16E-05
when in 0,000469 1 0,1538462 0,4 2,89E-05
i
i
i w
vote
vote +
=

54
Fig. 4. Context weight classification
In the table (Table 1), we note that the context "Hotel in" is found 17 times, in the corpus with the
learning examples. The context frequency is:
0,0039869
4264
17
=
=
in
Hotel
cf (10)
The context is found in 9 documents, but among these documents, only 4 documents come from
different sources. The document frequency for this context is:
0,4444445
9
4
=
=
in
Hotel
df (11)
The context "Hotel in" occurs 17 times in total, but only with 7 learning examples, the learning
example frequency is then:
0,5384616
13
7
=
=
in
Hotel
lef (12)
The context occurs 17 times with learning examples, but also 2 times with other phrases, the
inverse context frequency is:
5
,
8
2
17
=
=
in
Hotel
icf (13)
The weight of this context is then:
008110106
.
0
=
×
×
×
= in
Hotel
in
Hotel
in
Hotel
in
Hotel
in
Hotel icf
lef
df
cf
W (14)
The context "Hotel in" has the highest weight, which makes it more relevant for the recognition
of a capital name, compared to the other contexts.
We studied in addition, if the context number may change if we increase the size of the
analyzed corpus (the document number). In the table (Table 2), we present some statistics
associated to the ratio between the contexts number, and the corpus size. These numbers are
0
0,001
0,002
0,003
0,004
0,005
0,006
0,007
0,008
0,009
frequence
W
weight
Hotels in
Map of
trav
el to
Guide to
Places in
hotels in
map of
Hospital in
when in

55
found for the NE class "President", using 89 representations of president names as learning
examples:
"Jacques Chirac", "JACQUES CHIRAC", "Chirac", "Chirac", "CHIRAC", "Jacques
CHIRAC"; "Bush", "George Walker Bush", "George Bush", "GEORGE BUSH", "GEORGE W
BUSH", "GEORGE W. BUSH", "George W Bush", "BUSH", "George W. BUSH", "Geroge
BUSH", "Georges W.", "Georges W", "Nicolas Sarkozy", "NICOLAS SARKOZY", "Sarkozy",
"Sarko", "sarkozy", "Nicolas sarkozy", "Nicolas SARKOZY", "SARKOZY", "Jalal Talabani",
"Talabani", "TALABANI", "Jalal TALABANI", "JALAL TALABANI", "Zine El Abidine Ben
Ali", "Zine el Abidine", etc.
Table 2. Corpus size influence.
Document number Learning Example
occurrence number
Context number
(2 words left)
80 3003 1911
110 3637 2379
160 4699 2850
200 6894 3120
240 8592 3215
280 9260 3496
320 10362 3516
360 11184 3578
420 12218 3618
460 12981 3634
We have noticed that the context number becomes stable from a significant size corpus (420
records = 7 MB of text data).
Recall and precision are usually admitted approachs of measuring system performance in this
field. We suppose that Recall is the number of correct NE found by the system over total number
of correct NE in the document corpus. Precision is the number of correct NE found by the system
over total number of NE found in the corpus. We obtained the following value: for Precision 80,9
% and recall 69,8 %.
5.Conclusion
In this paper, the use of learning approach is suggested for the problem of NE recognition. The
goal is to uncover in a document corpus, NE that occurs frequently accompanied by contexts:
sequence of words, that are left or right of the NE. Contexts that occur with given learning
examples were first extracted from Web documents corpus. Different feature weighting measures
were examined to classify the contexts in order to identify the most pertinent contexts for the
recognition of a NE. This classification enables to derive a model for the recognition of a NE.

56
The same strategy can be applied to person names, company names, and many other types of
named entities in any language. Although, we should mention that we successfully applied this
technique to several named-entity types, in French and in English language.
One of the future works which we recommend is to discern and to measure similarity between
contexts. We can use this measurement to cluster similar contexts.
Since we have primarily applied our approach to the named entities problem, we can also
attempt additional concepts. In effect, the context usually contains enough information to identify
the instance as a concept member.
References
[1] M. Collins and Y. Singer, Unsupervised models for named entity classification, in Proceedings of the Joint
SIGDAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora, 1999, pp.
189–196,
[2] B. Daille, and E. Morin, Reconnaissance automatique des noms propres de la langue écrite: les récentes TAL,
Traitement automatique des langues, 2000, vol. 41, no
3, pp. 601-621.
[3] F. Denis, R. Gilleron, and F. Letouzey, Learning from positive and unlabeled examples. Elsevier. Theoretical
Computer Science, 2005, vol. 348, pp. 70 – 83
[4] O. Etzioni, M. Cafarella, D. Downey, S. Kok, A. Popescu, T. Shaked, S. Soderland, D. Weld, and A.Yates,
Unsupervised named-entity extraction from the web: An experimental study, Artificial Intelligence, 2005, vol. 65,
pp. 91–134.
[5] B. Favre, F. Béchet, and P. Nocéra, Robust Named Entity Extraction from Spoken Archives, in Proceedings of
HLT-EMNLP'05, pp. 491-498, Vancouver, Canada, October 2005.
[6] N. Fourour, and E.Morin, Apport du Web dans la reconnaissance des entités nommées. Revue québécoise de
linguistique, 2003, vol. 32, n° 1, pp. 41-60.
[7] N. Friburger, Linguistique et reconnaissance automatique des noms propres, Meta : journal des traducteurs,
2006, vol. 51, n° 4, pp. 637-650.
[8] C. Jacquemin, and C. Bush, Fouille du Web pour la collecte d’entités nommées, in Proceedings 8ème Conférence
Nationale sur le Traitement Automatique des Langues Naturelles (TALN 2000), pp. 187-196. Lausanne, 2000.
[9] C. Krstev, D. Vitas, D. Maurel, M. Tran, Multilingual ontology of proper name, in Proceedings of the Language
and Technology Conference, pp. 116–119, Poznan, Poland, 2005.
[10] G.S. Mann, Fine-grained proper noun anthologies for question answering, International Conference on
Computational Linguistics, COLING-02 on SEMANET: building and using semantic networks, 2002, Vol. 11,
pp. 1-7.
[11] C.Martineau, E. Tolone, and S. Voyatzi, Les Entités Nommées : usage et degrés de précision et de
désambiguïsation. In proceedings of 26th conference on Lexis and Grammar, pp. 105-112, Bonifacio, 2-6
October 2007.
[12] A. Mikheev, M. Moens, and C. Grover, Named Entity Recognition without Gazetteers, in Proceedings of
Conference of European, Chapter of the Association for Computational Linguistics, EACL '99, pp. 1-8,
University of Bergen, Bergen, Norway June 1999.
[13] D. Nadeau, P. D. Turney, and S. Matwin, Unsupervised named-entity recognition: Generating gazetteers and
resolving ambiguity, Lecture Notes in Computer Science, Springer, 2006, pp. 266–277, Berlin Heidelberg 2006.
[14] G Petasis, F Vichot, F Wolinski, G Paliouras, V. Karkaletsis, and C. D. Spyropoulos, Using machine learning to
maintain rule-based named-entity recognition and classification, in Proceedings of the 39th Annual Meeting on
Association for Computational Linguistics, pp. 426 – 43, Toulouse, France, 2001.
[15] J.R. Quinlan, C4.5: Programs for Machine Learning, Morgan Kaufmann, Los Altos, CA, 1993.
[16] S. K. Saha, S. Sarkar, and P. Mitra, Feature selection techniques for maximum entropy based biomedical named
entity recognition, Journal of Biomedical Informatics, Vol. 42, Issue 5, pp. 905-911, Elsevier, October 2009.
[17] G. Salton, A. Wong,, and C.S. Yang, A vector space model for information retrieval, Journal of the American
Society for Information Science, 1975; vol. 8(11), pp. 613-620.

Named Entity Recognition Using Web Document Corpus

More Related Content

What's hot (19)

Similar to Named Entity Recognition Using Web Document Corpus (20)

More from IJMIT JOURNAL (20)

Recently uploaded (20)

Named Entity Recognition Using Web Document Corpus