SEMI-SUPERVISED BOOTSTRAPPING APPROACH FOR NAMED ENTITY RECOGNITION

International Journal on Natural Language Computing (IJNLC) Vol. 4, No.5, October 2015
DOI: 10.5121/ijnlc.2015.4501 1
SEMI-SUPERVISED BOOTSTRAPPING
APPROACH FOR NAMED ENTITY
RECOGNITION
S. Thenmalar, J. Balaji and T.V. Geetha
Department of Computer Science and Engineering, Anna University, Chennai.
ABSTRACT
The aim of Named Entity Recognition (NER) is to identify references of named entities in unstructured
documents, and to classify them into pre-defined semantic categories. NER often aids from added
background knowledge in the form of gazetteers. However using such a collection does not deal with name
variants and cannot resolve ambiguities associated in identifying the entities in context and associating
them with predefined categories. We present a semi-supervised NER approach that starts with identifying
named entities with a small set of training data. Using the identified named entities, the word and the
context features are used to define the pattern. This pattern of each named entity category is used as a seed
pattern to identify the named entities in the test set. Pattern scoring and tuple value score enables the
generation of the new patterns to identify the named entity categories. We have evaluated the proposed
system for English language with the dataset of tagged (IEER) and untagged (CoNLL 2003) named entity
corpus and for Tamil language with the documents from the FIRE corpus and yield an average f-measure of
75% for both the languages.
KEYWORDS:
Named entity recognition, semi-supervised, pattern based bootstrapping, Tamil natural language
processing.
1. INTRODUCTION
In general, proper nouns are considered as named entities. The NER task was introduced during
the 6th
Message Understanding Conference (MUC) in 1996 [13], and in MUC- 7 [3] the initial
classification of named entities used the following categories and subcategories: Entity
(ENAMEX): person, organization, location, Time expression (TIMEX): date, time and Numeric
expression (NUMEX): money, percent. However, named entity tasks often include as named
entities expressions for date and time, names of sports and adventure activities, terms for
biological species and substances. The major challenge of named entity recognition is that of
tagging sequences of words that represent interesting entities, such as people, places, and
organizations. NER is a two-step process, the first step being the identification of proper nouns
which is the marking of the presence of a word or phrase as named entity (NE) in a given
sentence while the second step is its classification where the role of the identified NE is
determined.
NER was initially known as a significant component for Information Extraction (IE). NER has
now become vital for many other natural language processing based applications. The
identification and the semantic categories of Named entities are necessary before recognizing

2
relations between these entities [10, 11, 36]. NE’s play an important role in identifying
ontological concepts for populating ontologies [5, 12]. NEs convey the crucial information that
drives Information Retrieval (IR) and Question Answering (QA) systems [23, 35]. In more recent
times, important applications like news aggregation are usually centred on entities.
There are several approaches for identifying NER. The rule-based approach uses a set of rules
defined by human experts to extract entities. This model takes a set of patterns consisting of
grammatical, syntactic and orthographic features in combination with dictionaries. However,
manual creation of rules is labour intensive and costly and requires significant language as well as
domain expertise [33]. Moreover systems developed for one domain cannot be ported to another
domain. Therefore learning based approaches have been introduced for NER. Learning
algorithms can be defined as methods that use the features of training data and automatically
induce patterns for recognising similar information from unseen data. Learning algorithms can be
generally classified into three types: supervised learning, semi-supervised learning and
unsupervised learning. Supervised learning utilises only the labelled data to generate a model.
Semi-supervised learning aims to combine both the labelled data as well unlabelled data in
learning. Unsupervised learning aims to learn without any labelled data.
We present a semi-supervised pattern based bootstrapping approach to NER that automatically
identifies and classifies the entities. Our approach starts with the small set of tagged training data.
The tagged training data is used to identify the word and context features to define a five window
context pattern for each named entity category. We explore the representation of features used for
both English and Tamil languages to define the pattern. The identified patterns are used as seed
patterns. These seed patterns are used to identify the entities as an exact match in the test set. The
pattern scoring and the tuple value scoring decide the modification needed to generate new
patterns. The pattern score identifies which set of patterns are used for the next iteration. The
tuple value scoring of POS provides which set of tuple contributes to the named entity and
decides the window movement that is shift to the left or to the right and masks one tuple thus
generating of new patterns that is used to learn new context to identify Named entities.
The rest of the paper is organized as follows. Section 2 gives a brief description of related work
especially in the area of machine learning approaches to NER. Section 3 deals with the
bootstrapping approach to NER. Section 4 deals with evaluation and results while section 5 gives
a conclusion and discusses future work.
2. RELATED WORK
In a supervised learning approach an NER system takes training data and their features as input to
generate an extraction model, which is then used to identify similar objects in new data.
Supervised learning has been the most commonly used and the leading approach in the NER [28].
There are several widely used machine learning techniques for this task. Support Vector
Machines (SVM) [15, 7] a model is constructed that fits a hyperplane that best splits positive and
negative examples in the labelled data. The model characterizes the examples as points in space,
represented so that the positive and negative examples are separated by a clear gap that is as wide
as possible. New examples are represented into that same space during the application time and
expected to belong to a category based on which side of the gap they fall on.
Hidden Markov Model (HMM) [46, 44, 30] is a statistical Markov model in which the sequence
of states are hidden but can be predicted from a sequence of observations conveyed as a
probabilistic function of the states. The learning process in the context of NER concludes that an
HMM is based on the observed features and tags present in the training data. The model generates

3
a mapping with certain probability that can predict a sequence of states. Conditional Random
Fields (CRF) [17, 1] is a probabilistic model which avoids certain assumptions about the input
and output sequence distributions of HMM. The other broadly used machine learning techniques
for detecting NER where Perceptron algorithms [18], Naïve Bayes [27], Decision Trees [9] and
Maximum Entropy model [2]. For Tamil language named entities are identified by using
Expectation Maximisation and the CRF model [29, 41]. The identification for named entity using
CRF for Tamil language [41] describes the characteristics feature and handles morphological
inflections to represent the training model. In our approach, we consider the morphological
suffices as an added feature to represent the pattern for the named entity categories. The
supervised learning method depends on the large set of training data, which has to be annotated
manually.
Unsupervised learning methods recognize named entity based on unlabelled data. The
unsupervised learning methods basically use clustering techniques, distribution statistics and
similarity based functions. The recognition of various types of named entities in the open domain
that could be useful in IE [8]. The sequence of capitalised words that are likely to be named
entities are extracted and the search queries using the sequence of words are created with Hearst
patterns [14]. The hypernyms were extracted and clustered. The entities are looked up in
WordNet and are labelled by the top level concepts observed in the WordNet. The named entities
are lexicalised as multi-word in which co-occuring terms occur more frequently [6]. They have
identified possible n-grams entity from the corpus based on the mutual information measures and
the frequency of words and grouped similar named entities using clustering algorithm. The
complex named entities in the Web data is identified using clustering algorithm [45]. The named
entities are labelled based on the similarity using vector similarity model [4]. In essence, the
entity names and their types are described as vectors with the specified features. In Semantic
Concept Mapping, with the known list of candidate entity names and labels are denoted as
WordNet synsets [19]. The Lin’s similarity function describes the type of entity name [25].
We explore the semi-supervised learning method for NER. The semi-supervised method uses the
small number of labelled data to learn and tag a large set of unlabelled data. The self-training
algorithm is used for detecting named entity and voted co-training algorithm is used for
classifying the named entity [21]. The self-training algorithm that selects the unlabelled instances
for the Naïve Bayes classifier [43]. The gene name is recognized by bootstrapping approach [42].
The abstract and the list of genes are available for the set of articles. The gene mentions in the
abstract were annotated with the list of gene available for the article. The HMM-based tagger is
trained with the annotated data. The lists of entities stated in the document are not usually
available in general named entity recognition. In another approach the named entities having
Heidelberg Named Entity Resource are classified based on the Wikipedia category using
bootstrapping [20]. The inconsistency in classification might be undetermined by placing the
articles to the specified category. The features of POS and syntactic structure of the document are
used for the semi-supervised learning algorithm; a Self-Training algorithm is used to recognize
the named entity [34]. The features used determine the entity boundary and pattern extraction.
However in our approach we use the POS feature and the context information of the word to
represent the pattern. The CRF with feature induction is used for detecting Hindi NER [24]. The
features induction includes word based features, character n-grams, word prefix and suffix and 24
gazetteers. However we explore detecting named entity without gazetteers.
NER problem is considered to be solved for languages such as English but still remains a
challenge for resource scarce languages. Moreover NER of informal texts such as tweets and
blogs still remains a challenging problem [22]. The issues in handling NER tasks for Indian
languages have been described in the survey of Named Entity Recognition [16]. The issues

4
discussed are the major feature commonly followed by NER systems is the capitalization of
words. However, the capitalization of word for the named entities is not represented in the Indian
languages. The gazetteers of named entities are unavailable for Indian languages. Additionally,
spelling variations are common in Indian languages. For example in Tamil language: ராசா
(Rasa) - ராஜா (Raja): Person named Raja, புதுேசr (puducherri) - புதுச்ேசr
(pudhuccherri) : Place named Puducherry. The lack of labelled data is the added issue to resource
scarce and morphological rich languages. In this work, we use pattern representation of
bootstrapping approach which requires only a small set of feature tagged seed samples to learn
the context and the features that detect the named entity and its category.
3. BOOTSTRAPPING PROCESS FOR NAMED ENTITY
RECOGNITION
Figure 1 shows the overall bootstrapping process for NER. The proposed approach starts with the
small set of training examples. The training set of documents is manually annotated with the
named entity categories. The annotated training set is pre-processed by identifying the features of
the word. We make use of the context of the word to define the pattern. The patterns associated
with each category of named entity are identified and used as seed patterns. The test data is
processed by matching the features of the word with the pattern. If exact match occurs then the
named entity category is identified. Up to this point we have identified and categorized named
entities that have features exactly similar to the seed set initially given. However we need to
generate new patterns by learning new contexts where these named entities can occur. For this
purpose the patterns are scored to identify which patterns can be used for further iteration. The
new patterns have also been designed to identify named entity chunks even though the initial seed
patterns are associated with only a single word. The new pattern is generated by right or left shift
of the window for each pattern depending on the tuple value score. The new pattern is given as
input to the tagged test documents and the instances of the named entity categories are identified.
The process of pattern scoring, tuple value scoring and window shifting to learn new patterns is
continued until no new patterns are generated or all the test data is labelled. The feature set used
is described in the next section.

5
Figure 1 Overall bootstrapping process for Named Entity Recognition
3.1 Feature set
A perceptron based recognizer for identifying named entities uses nonlocal dependencies and
external information as features [31]. A supervised learning method with CRF for detecting NER
uses local knowledge features, external knowledge features and the non-local dependencies [38].
They have discussed that the system when using the local knowledge feature performs poorly
when using the single token and the maximum observation of named entities is shown when
using the sliding window of 3 token. However, when considering a three window context, the
ambiguity of the type of named entity occurs. The features used also depend on the language
under consideration. In order to overcome the above issue we go for a five window context (wi-
2,wi-1,wi,wi+1,wi+2)
The commonly used features for NER systems were part-of speech tags, shallow parsing and the
gazetteers. Part-of-speech (POS) tags are commonly used feature in NER but this feature is not
considered for NER Systems [31, 26]. In this work, we use POS tag and semantic constraints
obtained from UNL KB [40] that are associated with each word along with the five window
context as common feature for both English and Tamil languages. Thus each word is attached
with POS tag and semantic constraint to form the feature set. Using these features we describe the
patterns.
Tagged
document
Seed Pattern Generation
Matching
Processed document
corpus (test
Identification of
Named Entities
Scoring
MaskinGeneration of
new patterns

6
3.2 Seed pattern Generation
Tagging the words in the English documents with POS tag is carried out using the Stanford parser
and the POS tagging of the words in the Tamil documents is carried out using morphological
analyser. We manually label the named entities in a small set of tagged data. This data is
considered as training data and is used for identifying the patterns. The patterns capture both the
sequence of tokens that identify a potential named entity and the information from the right & left
context where it occurs. We consider the five window context (wi-2,wi-1,wi,wi+1,wi+2) to represent
the pattern. We have defined two types of patterns, one type for English language where the POS
tag and semantic constraint (SC) are the features associated with each word is given below. In the
case the pattern type of Tamil language we also use morphological suffix (MS) which implicitly
conveys case information for nouns as an additional feature is given below. This difference in the
type of patterns essentially caters to two languages having different characteristics such as fixed
word order of English language and partially free word order of Tamil language. The use of word
based semantic constraint (SC) allows the context of the named entity is to be semantically
described to enable proper classification. Thus each word in the context window (wi-2,wi-
1,wi,wi+1,wi+2) consists two tuples in the case of English language and three tuples in the case of
Tamil language.
Pattern Type for English language
POS,SC(wi-2);POS,SC(wi-1);POS(wi),SC@Named Entity Type; POS,SC(wi+1);POS,SC(wi+2)
Pattern Type for Tamil language
POS,MS& SC(wi-2);POS,MS&SC(wi-1);POS,MS & SC (wi) @Named Entity Type;
POS,MS&SC(wi+1);POS,MS&SC(wi+2)
The Named Entity Types used here are defined in the the classification of MUC-7 namely person,
organization, location, date, time, money and percentage.
Example Pattern Types for English language
Person
VBG,icl>person; IN,aoj>thing; NNP,iof>person@person; IN,aoj>thing; PRP,icl>female person
Location
TO,aoj>thing; VB,agt>thing,obj>thing; NP,iof>country@location; NN,icl>area; NNS,icl>action
Organization
IN,obj>thing; IN,aoj>thing; nnp,icl>organization@organization; JJ,aoj>thing; NN,icl>facilities
Date
NN,icl>organization; IN,aoj>thing; NNP,icl<date; CD,None@date; VBD,obj>thing;
NNS,aoj>thing
Time
NN,icl>action; VBD,None; NN,icl>time@time; IN,aoj>thing; DT,None
Money
NN,icl>deal; IN,aoj>thing; $CD,icl>money@money; IN,aoj>thing; NN,icl>reduce

7
Percent
VB;agt>thing,obj>thing; PRP,None; CD,None;NN,icl>ratio@percent; NN,agt>thing;
IN,aoj>thing
Example Pattern Type for Tamil language
Person
Entity,அ,icl>region; Noun,None,icl>person; Entity,None,iof>person@person;
Adjective,ப்,icl>help; Adjective,None,aoj>thing
Location
Noun,இன்,icl<weather; Noun,ஆக,icl<abstract thing; Entity,None,iof>place@location;
Noun,இல்,icl<area; Noun,None,icl>calculate(agt>thing,obj>thing)
Organization
Noun,None,icl>person; Noun,உடன்,icl>act; Noun,None,icl>organization@organization;
Noun,கள்,icl>person; Verb,None,icl>action
Date
Adjective,None,mod<thing; DateTime,None,icl>period; DateTime,None,icl>month,
charNumbers,ஆம்@date; DateTime,None,aoj>thing; Noun,உக்கு+ச்,icl>facilities
Money
Pronoun,None,icl>person; Noun,None,aoj>thing; charNumbers,None,Noun,ஐ,
icl>currency@money;Adverb,None,icl>action; Noun,ஆ ,aoj>thing
Time
Noun,None,icl>morning; charNumbers,None,Noun,க்கு,icl>time; Noun,None, icl>workship;
Verb,உம்,icl>action
Percent
Noun,கள்,icl>person; Noun,க் + ைகயில்,icl>action; charNumbers,None,Noun, icl>ratio
@percent; Noun,ஆக,icl>change; Verb,None,agt>thing,gol>person,obj>thing
3.3 Matching
For a given test data we POS tag the words using Stanford parser in case of English language and
use a morphological analyser [39] in the case of Tamil language for POS and Morphological
suffix tagging. The root words are then used to obtain the corresponding semantic constraints
from the UNL KB [40]. We check for the matching of seed patterns with the annotated sentences
of the documents. Although the pattern consists of a five word window (wi-2,wi-1,wi,wi+1,wi+2),
actual exact matching is carried out with only the middle word wi of the pattern. If there is a
match the corresponding classes are labelled for the exactly matched patterns. Named entities that
are not handled by the exact match are processed through partial matching. Partial match is
carried out after selecting the pattern to be modified during iteration in the next cycle. Once a
pattern is identified, tuple scoring detects which tuple contributes the most to the particular entity
type.
3.4 Generation of new patterns
The first step in new pattern generation is to find the most frequently occurring pattern for each
class of named entity indicated by Pattern Score. The next step is to find alternate values for POS

8
tags that can occur at position k in the context window of pattern Pj keeping all other tuple values
the same by finding the tuple value of POS tag with minimum score at position k which is then
considered for masking.
3.4.1 Pattern Score
In this work we use the Basilisk algorithm for calculating the pattern scoring metric RlogF metric
[32]. The extraction pattern is scored using the following formula:
(1)
Where Fj is the number of identified named entities by pattern Pj corresponding to a particular
type of named entity, nj is the total number of patterns identified. The pattern score identifies the
pattern for the particular type of named entity to be chosen for modification to form the new
pattern.
3.4.2 Tuple value Score
The tuple scoring basically depends on the tuple corresponding to the POS tag of the words in the
context window. This scoring essentially evaluates which POS tag value of which word in the
context window of the specific pattern is strongly associated with a particular pattern.
Let us consider a pattern Pj corresponding to a particular type of Named entity. The pattern P
associated with Named entity has four words (wi-2,wi-1,wi,wi+1,wi+2)in the 5 word context window
and each of these words is associated with a POS values (posi-2, posi-1, posi+1, posi+2). The
maximum score of the POS value indicates that this POS value in this position contributes the
most to the pattern Pj and is given as
where k=i-2,i-1,i+1,i+2 (2)
Here, tvPOSk corresponds to tuple value of POS tuple posk.
f(tvPOSk, Pj) is the number of times this particular POS value at position k occurs with
pattern Pj.
f(Pj) is the frequency of pattern Pj and
f(tvPOSk) is the total frequency of this tuple value at position k.
3.4.3 Methods for New Pattern Generation
The first method of new pattern generation is the replacement of POS value in the appropriate
position k. The POS tuple value with minimum score at position k is masked. The new pattern is
generated by replacing the tuple value of the original pattern by the new tuple value that occurs
the most frequently at position k in the test data.
The next method of new pattern generation is carried out by shifting the context window to the
left or right of the Wi depending on the frequency of occurrence of POS pair of wi, wi+1 or wi-1, wi.
Depending on a higher POS pair frequency score a new 5 word pattern is generated where either
wi-1 or w i+1 becomes the new wi. This method of new pattern generation is possible since Named
Entities are often associated with POS tags that frequently tend to co-occur together.

9
New patterns are also generated by chunking words to form phrasal named entities. For this
purpose again we use the POS pair frequency score as shown in Eq. 1, but in addition we check
whether each of the words associated with POS pair have the same semantic constraints. In case
the POS pair frequency is above a threshold and have same semantic constraints they are chunked
as a single Named entity and considered as wi for the next iteration. This unique way of forming
patterns for chunked words forming Named Entities is possible because these chunks are often
associated with similar semantic properties. However the two languages we considered needed to
be tackled differently during the chunking process.
In the case of English language, the POS pair frequency score and semantic constraints alone
decide chunking. However in the case of Tamil language, in addition to the above features,
morphological suffix should not be associated with Wi in case of pair Wi, Wi+1 or Wi-1 in case
pair Wi-1, Wi. In case such morphological suffixes exist chunking is not carried out even though
the semantic constraints between the pair matches.
4. EVALUATION
We have tested the performance of the system on IEER dataset, the tagged corpus which contains
the Newswire development test data for the NIST 1999 IE-ER Evaluation. There were totally
3174 named entities. We have taken two different seed patterns that commonly occur for each
named entity class (MUC-7). Iterations are carried out until no changes occur in the patterns. The
performance of the system is evaluated using precision, recall, and f-measure metrics. Here, recall
is defined as
Precision is defined as
F-measure is the weighted harmonic mean of precision and recall and is given as
The performance of NER system using IEER data set is given in Table 1, where the system
identifies the named entities with the average precision of 83% and recall of 92%.

10
Named Entities precision recall F-measure
person 84.72 93.29 88.8
Location 82.1 92.9 87.17
organization 83.08 92.19 87.4
Date 83.3 92.83 87.81
Time 81.53 91.37 86.17
Money 81.55 91.3 86.15
Percent 81.48 91.67 86.28
Table 1 Performance of Bootstrapping system using IEER dataset
We have tested the performance of the system by using the CoNLL 2003 [37] data RCV1
(Reuters Corpus Volume 1). The Reuters corpus consists of news articles between August 1996
and August 1997. The training set was taken from the files representing the end of august 1996.
For test set the files were from December 1996. There were totally 5648 Named Entities. The
training data is processed and tagged with POS tags and semantic constraints. We extract two
different seed patterns for each named entity class (MUC-7) from the training set. With the seed
patterns, we used the test set to identify exact match and iterate the process to generate new
patterns for the named entity class. The performance is shown in Table 2, where the system yields
the average precision of 80% and recall of 89%.
Table 2 Performance of our system for CoNLL 2003
For Tamil language we performed experiments on documents of the FIRE (Forum of
Information Retrieval Evaluation) Tamil corpus extracted from newspapers such as BBC,
Dinamani and Dinamalar. We have considered 50000 documents and tagged with the
appropriate features such as POS, Morphological suffix and UNL Semantic constraint. We
have taken 4000 tagged documents for training set and extracted the most frequently
occurring two different example patterns for each named entity class (MUC-7). The
performance is shown in Table 3, where the system produces the average precision of 79%
and recall of 88%.
person 84.34 90.46 87.29
location 75.16 90.01 81.92
date 75.15 86.3 80.34
time 72.02 86.56 78.62
money 73.02 86.42 79.16
percent 74.42 85.88 79.74
Table 3 Performance of our system in FIRE corpus.
person 82.82 93.39 87.79
location 81.62 90.52 85.84
date 79.54 87.5 83.33
time 75.84 89.5 82.11
money 71.05 81 75.7
percent 72.8 83 77.57

11
We also compared our bootstrapping with the baseline approach [34]. The Baseline system uses
Reuters corpus and considers accident documents. The total number of named entities
corresponding to the date and location are 246 and 596. The Baseline system uses 15 and 36 seed
patterns for date and location entities and extracted 18 and 68 patterns. Our system uses 2 seed
patterns for each named entities and we have learned 12 and 35 patterns of those named entities.
The comparison is shown in figure 2, the missing of syntactic structure information in the
baseline system yields less F-measure whereas our approach makes use of the larger contextual
window and word based semantic information to learn the patterns.
Figure 2. Performance of Baseline system and our Bootstrapping approach
We also shown in the Table 4 and Table 5 the number of patterns that we have learnt in the
English corpus and Tamil corpus for each named entities using each method (replacement of POS
and shifting window).
Number of New Patterns Learnt - CoNLL 2003
Person Location Organization Date Time Money Percent
Replacement of POS 26 32 28 17 15 12 8
Shifting the window 38 44 35 24 22 18 14
Table 4 Number of patterns learnt from the CoNLL 2003 English Corpus
Number of New Patterns Learnt - FIRE
Person Location Organization Date Time Money Percent
Replacement of POS 54 62 58 35 38 18 12
Shifting the window 75 83 77 48 42 24 22
Table 5 Number of patterns learnt from the FIRE Tamil Corpus

12
5. CONCLUSION
This paper describes a new pattern based semi-supervised bootstrapping for identifying and
classifying Named Entities. The method does not use any Gazetteer but instead uses POS
information and word based semantic constraints and gives an average f-measure of 75 % for
both the languages. This essentially ensures that the patterns are feature based enabling tagging
of hitherto unseen Named Entities. This method can be further enhanced by considering more
domain specific corpora and by trying for domain specific Named Entity categories. Future work
includes testing this method for other languages to study how the methodology needs to adapted.
Pattern definition and chunking strategies are possible aspects that need to be modified.
REFERENCES
[1] Arnold, A. Nallapati, R. and Cohen, W. 2008. Exploiting Feature Hierarchy for Transfer Learning in
Named Entity. Journal of Computational Linguistics, 245-53
[2] Chieu, H. and Ng, H. 2003. Named Entity Recognition with a Maximum Entropy Approach. In
Proceedings of the Seventh Conference on Natural Language Learning (CoNLL2003). 160-163.
[3] Chinchor, N. 1998. MUC-7 Named Entity Task Definition Version 3.5. In Proceedings of the 7th
Message Understanding Conference (MUC-7).
[4] Cimiano, P. and Völker, J. 2005. Towards Large-scale, Open-domain and Ontology-based Named
Entity Classification. In Proceedings of the International Conference on Recent Advances in Natural
Language Processing (RANLP’05). 166-172.
[5] Cimiano, Philipp. 2006. Ontology Learning and Population from Text: Algorithms, Evaluation and
Applications. 1st ed. Springer.
[6] Da Silva, J., Kozareva, Z. and Noncheva, V. 2004. Extracting Named Entities: a Statistical Approach.
In Proceedings of the XIème Conférence sur le Traitement des Langues Naturelles (TALN). 347-351.
[7] Ekbal, Asif. and Bandyopadhyay, Sivaji. 2010. Named Entity Recognition using Support Vector
Machine: a Language Independent Approach. International Journal of Electrical and Electronics
Engineering. 4(2), 155-170.
[8] Evans, R. 2003. A Framework for Named Entity Recognition in the Open Domain. In Proceedings of
Recent Advances in Natural Language Processing (RANLP2003). 137-144.
[9] Finkel, J. and Manning, C. 2009. Nested Named Entity Recognition. In Proceedings of the 2009
Conference on Empirical Methods in Natural Language Processing. 141-150.
[10] Giuliano, C., Lavelli, A. and Romano, L. 2006. Exploiting Shallow Linguistic Information for
Relation Extraction from Biomedical Literature. In proceedings of the 11th Conference of the
European Chapter of the Association for Computational Linguistics (EACL). 401-408.
[11]Giuliano, C., Lavelli, A. and Romano, L. 2007. Relation Extraction and the Influence of Automatic
Named Entity Recognition. ACM Transactions on Speech and Language Processing (TSLP). 5(1). 1-
26.
[12] Giuliano, C. and Gliozzo, A. 2008. Instance-Based Ontology Population Exploiting Named-Entity
Substitution. In Proceedings of the 22nd International Conference on Computational Linguistics. 265-
272.
[13] Grishman, Ralph. and Sundheim, Beth. 1996. Message understanding conference-6: a brief history. In
Proceedings of the 16th conference on Computational linguistics. Volume 1, 466–471.
[14] Hearst, M. 1992. Automatic Acquisition of Hyponyms from Large Text Corpora. In Proceedings of
the 14th International Conference on Computational Linguistics. 539-545.
[15] Isozaki, H. and Kazawa, H. 2002. Efficient Support Vectors Classifiers for Named Entity
Recognition. In Proceedings of the 19th International Conference on Computational Linguistics. 1-7.
[16] Kaur, Darvinder. and Gupta, Vishal. 2010. A survey of Named Entity Recognition in English and
other Indian Languages, IJCSI International Journal of Computer Science Issues. 7(6), 239-245.
[17] Kazama, J. and Torisawa, K. 2007a. Exploiting Wikipedia as External Knowledge for Named Entity
Recognition. In Proceedings of the Joint Conference on Empirical Methods in Natural Language
Processing and Computational Natural Language Learning. 698-707.

13
[18] Kazama, J. and Torisawa, K. 2007b. A New Perceptron Algorithm for Sequence Labeling with Non-
local Features. In Proceedings of the 2007 Joint Conference on Empirical Methods in Natural
Language Processing and ComputationalNatural Language Learning. 315-324.
[19] Kliegr, T., Chandramouli, K., Nemrava, J., Svatek, V. and Izquierdo, E. 2008. Combining Image
Captions and Visual Analysis for Image Concept Classification. In Proceedings of the 9th
International Workshop on Multimedia Data Mining, at the 2008 International Conferenc on
Knowledge Discovery and Data Mining. 8-17.
[20] Knopp, J. 2011. Extending a Multilingual Lexical Resource by Bootstrapping Named Entity
Classification using Wikipedia’s Category System. In Proceedings of the 5th International Workshop
On Cross Lingual Information Access, at IJCNLP2011 . 35-43.
[21] Kozareva, Zornitsa. Bonev, Boyan. and Montoyo, Andres. 2005. Self-training and Co-training
Applied to Spanish Named Entity Recognition. MICAI 2005, Advances in Artificial Intelligence, 770-
779.
[22] Kucuk, Dilek. Jacquet, Guillaume. and Steinberger, Ralf. 2014. Named Entity Recognition on Turkish
Tweets. In Proceedings of the Ninth International Conference on Language Resources and Evaluation
(LREC'14). 26-31.
[23] Lee, C., Hwang, Y. and Jang, M. 2007. Fine-grained Named Entity Recognition and Relation
Extraction for Question Answering. In Proceedings of the 30th annual international ACM SIGIR
conference on Research and development in information retrieval SIGIR'07. 799-800.
[24] Li, W. and McCallum, A. 2004.Rapid Development of Hindi Named Entity Recognition using
Conditional Random Fields and Feature Induction. ACM Transactions on Asian Language
Information Processing, 2(3), 290-294.
[25] Lin, D. 1998. An Information-theoretic Definition of Similarity. In Proceedings of the Fifteenth
International Conference on Machine learning (ICML'98). 296-304.
[26] Lin, Dekang. and Wu, Xiaoyun. 2009. Phrase Clustering for Discriminative Learning. In the
Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International
Joint Conference on Natural Language Processing of the AFNLP. 1030-1038.
[27] Mohit, B. and Hwa, R. 2005. Syntax-based Semi-supervised Named Entity Tagging. In Proceedings
of the ACL Interactive Poster and Demonstration Sessions. 57-60.
[28] Nadeau, D. 2007. A Survey on Named Entity Recognition and Classification. Linguisticae
Investigationes. 30(1), 1-26.
[29] Pandian, S., Geetha, T. V. and Krishna. 2007. Named Entity Recognition in Tamil using Context-cues
and the E-M Algorithm. In Proceedings of the International Joint Conference on Artificial
Intelligence. 1951-1958.
[30] Ponomareva, N., Pla, F., Molina, A. and Rosso, P. 2007. Biomedical Named Entity Recognition: a
Poor Knowledge HMM - based Approach. In Proceedings of the 12th international conference on
Applications of Natural Language to Information Systems (NLDB). 382-387.
[31] Ratinov, Lev. and Roth, Dan. 2009, Design Challenges and Misconceptions in Named Entity
Recognition. In the Proceedings of the Thirteenth Conference on Computational Natural Language
Learning CoNLL’09. 147-155.
[32] Rillof, E. 1996. An Empirical Study of Automated Dictionary Construction for Information Extraction
in Three Domains. Al Journal, vol. 85. 101-134.
[33] Sarawagi, S. 2007. Information Extraction. Foundations and Trends in Databases , Vol. 1(3), pp.261-
377.
[34] Sari, Yunita. Hassan, Mohd F. and Zamin, Norshuhani. 2010. Rule-based pattern extractor and named
entity recognition: A hybrid approach. In Proceedings of International Symposium on Information
Technology, ITSim. 563-568.
[35] Srihari, R. and Peterson, E. 2008. Named Entity Recognition for Improving Retrieval and Translation.
In Proceedings of the 11th International Conference on Asian Digital Libraries. 404-05.
[36] Thahir, M., Estevam, H. and Mitchell, T. 2011. Discovering Relations between Noun Categories. In
Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP) .
1447-1455.
[37] Tjong Kim Sang, Erik F. and De Meulder, Fien. 2003. Introduction to the CoNLL-2003 Shared
Task:Language-Independent Named Entity Recognition. In the Proceedings of CoNLL-2003. 142-
147.
[38] Tkachenko, Maksim. and Simanovsky, Andrey. 2012. Named entity recognition: Exploring features.
In the Proceedings of KONVENS 2012. 118-127.

14
[39] Umamaheswari E, Karthika Ranganathan, Geetha T.V, Ranjani Parthasarathi, Madhan K. 2011.
Enhancement of morphological analyzer with compound, numeral and colloquial word handler, In the
Proceedings of International Conference on Natural Language Processing (ICON), 177-186.
[40] UNDL 2005,UNDL Foundation, Geneva, Switzerland, 2005. www.undl.org.
[41] Vijayakrishna R and Sobha L. 2008. Domain focused named entity recognizer for tamil using
conditional random fields. In Proceedings of the IJCNLP-08 Workshop on NER for South and South
East Asian Languages, 93–100.
[42] Vlachos, A. and Gasperin, C. 2006. Bootstrapping and Evaluating Named Entity Recognition in the
Biomedical Domain. In Proceedings of the HLT-NAACL BioNLP Workshop on Linking Natural
Language and Biology . 138-45.
[43] Wang, Bin. Spencer, Bruce. Ling, Charles X. and Zhang, Harry. 2008. Semi-supervised Self-training
for Sentence Subjectivity Classification. in Advances in Artificial Intelligence. 344-355.
[44] Zhou, G. and Su, J. 2004. Exploring Deep Knowledge Resources in Biomedical Name Recognition. In
Proceedings of the COLING2004 International Workshop on Natural Language Processing in
Biomedicine and its Applications. 99-102.
[45] Downey, D, Broadhead, M & Etzioni, O 2007, Locating Complex Named Entities in Web Text,
Proceedings of the 20th International Joint Conference on Artificial Intelligence, 2733-2739.
[46] Morwal, S, Jahan, N and Chopra, D, 2012, Named Entity Recognition using Hidden Markov Model,
International Journal on Natural Language Computing (IJNLC) Vol. 1, No.4, December 2012

SEMI-SUPERVISED BOOTSTRAPPING APPROACH FOR NAMED ENTITY RECOGNITION

More Related Content

What's hot (20)

Similar to SEMI-SUPERVISED BOOTSTRAPPING APPROACH FOR NAMED ENTITY RECOGNITION (20)

More from kevig (20)

Recently uploaded (20)

SEMI-SUPERVISED BOOTSTRAPPING APPROACH FOR NAMED ENTITY RECOGNITION