A STUDY ON NLP APPLICATIONS AND AMBIGUITY

PROBLEMS
Anuj Yadav

ITM UNIVERSITY GWALIOR

ABSTRACT

Natural language processing (NLP) has been considered as one of the important area in Artificial
Intelligence. However, the progress made in natural language processing is quite slow, compared to other
areas. The aim of this study is to conduct a systematic literature review for identifying the most prominent
applications, techniques and challenging issues in NLP applications. To conduct this review, I had screened
587 retrieved papers from major databases such as SCOPUS and IEEE Explore, and also from Google
search engine. In searching relevant papers search keywords such as "natural language processing, NLP
applications, and complexity of NLP applications" had been used. However, to focus to the scope of the
study 503 papers were excluded. Only the most prominent NLP applications namely information extraction,
question answering system and automated text summarization were chosen to be reviewed. It is obvious
that the challenging issue in NLP is the complexity of the natural language itself, which is the ambiguity
problems that occur in various level of the language. This paper also aims at addressing ambiguity
problems which occur at lexical and structural levels and significance techniques or approaches for solving
the problems. Finally, the paper briefly discuss the future of NLP.
Keywords: Natural language processing, NLP applications, Ambiguity in NLP

1. INTRODUCTION by humans. The foundation of NLP is basically lies

on many disciplines, such as linguistics,
Information is normally stored in text documents. mathematics, computer sciences, psychology, and
These text documents can be found on personal so on.
desktop computers, intranets and on the Web.
Research work in NLP has been ongoing since
Valuable knowledge is normally embedded inside
more than six decades. Alan Turing’s article
unstructured texts. The Web has been considered
“Computing Machinery and Intelligence” published
as the world’s largest repository of knowledge, and
in 1950s has been considered as the earliest work in
it is being constantly augmented and maintained by
NLP. Later in 1954, the work of Georgetown
millions of people around the world. However, it is
University and IBM on translating 60 Russian
not in the form of a database from which records
sentences into English language was considered
and fields are easily manipulated and understood by
among the earliest successful in NLP. In the year of
computers, but in natural language texts which are
1978-1970, Terry Winograd, an American professor
intended for human reading. In spite of the promise
of computer science at Stanford University showed
of the semantic web, the use of English language
a successful computer program that can understand
and other natural language texts will continue to be
a natural language. The program was named as
a major medium for communication, knowledge
SHRDLU, developed using Lisp programming
accumulation, and information distribution [1]. This
language. During the 70’s NLP are based on
requires the study of NLP; a field of computer
conceptual ontologies and real world information.
science and linguistics, concerning with the
This field is moving rapidly and much work has
interaction between computers and a natural
been conducted in the last 10 years. Despite the
language. The term natural language is normally
goal of NLP’s work far from being completely
used to distinguish human languages such as Arabic
successful, significant positive outcomes has been
or English from formal languages such as C++, C,
shown in some research work. [2], [3], [4],[5] and
Java or XML. The ultimate goal of NLP researchers
[6]. The aim of this paper is to survey the most
is to create a software program that enables
prominent applications and techniques in NLP.
computers to understand or generate language used
Despite a number of survey papers on NLP have
been published so far, to the best of my knowledge, 3. RESULTS AND ANALYSIS
none of them have discussed about the most
NLP encompass both text and speech, however,
prominent NLP applications and approaches in
work on speech processing has evolved into
solving ambiguity problems comprehensively. This
separate fields. When NLP term is used, it normally
paper is organized as the following. Section 2
refers to natural language generation (NLG) and
presents research questions and methods used,
natural language understanding (NLU). NLG
section 3 presents results and analysis, section 4
involves with some form of computerized data into
address the future of NLP, and the paper is
natural language, rather than the other way around.
concluded in section 5.
However NLG is different than techniques for
2. RESEARCH AND METHODS ‘report generation’, ‘document generation’, ‘mail
merging’ and so on. These techniques are simply
This review is guided by the following 4
plugging a fixed data structure such as a table of
research questions:
numbers or a list of names into a template in order
RQ1: What are the different between natural to produce complete documents. On the other hand,
language understanding and natural language NLG uses some level of underlying linguistic
generation? representations of texts, to make sure that the
generated text is grammatically correct and fluent.
RQ2: What are the most prominent NLP
Most NLG systems include a syntactic reliazer to
applications and their techniques?
ensure that grammatical rules such as subject-verb
RQ3: What are the challenging issues in NLP? agreement are obeyed, and text planner to decide
how to arrange sentences, paragraph, and other
RQ4: What are approaches and techniques used to parts coherently. A machine translation system is
resolve ambiguity problems? the most known application of NLG.. The system
The first step of this systematic literature review is analyzes texts from a source language into a
screening all 587 retrieved documents. The process grammatical or conceptual representation, and then
of screening includes download the papers from the generates corresponding texts in the target language
major databases such as SCOPUS, IEEE Explore, [7]. On the other hand NLU is a system that
Google Search, and read their abstracts. Three computes the meaning representation, essentially
main keywords for search were used: natural restricting the discussion to the domain of
language processing, NLP applications, and computational linguistic [8]. NLU is independent
complexity of NLP applications. During this from speech recognition [9]. However, the
screening process 300 papers were excluded combination of the two may produce a powerful
because of poor presentation. Then 287 papers were human-computer interaction system. When
reviewed and evaluated. This process excluded 203 combined with NLU, speech recognition transcribes
papers. Only 84 papers were selected to be included an acoustic signal into a text. Then the text is
in this paper. The selection was made based on interpreted by an understanding component to
three criteria: relevant to the research questions, extract the meaning. The connection between
well written, and had been cited by other speech recognition analysis and natural language
researchers. Classifications of the selected papers understanding is illustrated in Figure 1.
are shown in Table 1.
Table 1: Classifications of extracted papers
Topic Sub-Topics References
NLP Application Information [11] to [25]
Extraction
Question Answering [26] to [32]
System

Automated Text [33] to [43]

Summarization
Challenging Issues Structural Ambiguity [45] to [64]
in NLP Figure 1. The diagram shows the connection between
Lexical /Word Sense [65] to [84]
speech recognition analysis and natural language
Ambiguity
understanding.
 NLP Applications information collected in the previous components is
Applications of NLP have been included analyzed to determine the relationships between
in many study fields, such as natural language user entities.
interface, automated text summarization,
information extraction, machine translation,
questions answering system, speech recognition,
text mining, and document retrieval and so on.
However, in this review the focus is given on
techniques used in information extraction,
automated text summarization, and questions and
answering system. These applications are normally
used as fundamental approaches for other types of
NLP applications. For example, information
extraction is a fundamental task in text mining data Figure 1. The diagram shows a database can be
analysis and text mining applications. populated through information extraction technique. The
populated data is then mined to extract the hidden
Information Extraction knowledge. The knowledge can be visualized through
Information Extraction (IE) is defined as visualization approach.
one of NLP's tasks for recognizing and extracting
instances of a particular pre-specified class of The major task in IE is known as named entity
entities, relationships, and events in natural recognition (NER). A process of named entity
language texts. recognition refers to the combined task of finding
spans of text that constitute proper names and then
The IE field has been initiated by DARPA’s MUC classifying the entities referred to according to their
program (Message Understanding Conference) in types. It is an important task in many NLP
1987. MUC has originally defined IE as the task of applications such as information retrieval, machine
(1) extracting specific, well-defined types of translation, and question answering systems. The
information from the text of homogeneous sets of most studied entity types are categorized into
documents in restricted domains and (2) fill pre- “proper names”, “names of persons”, “locations”,
defined form slots or templates with the extracted and “organizations”. In general, early research
information. The process of information extraction works of NER focus on recognizing “proper
is also called text analysis. It turns the unstructured names”. The “location” type can, in turn, be divided
information embedded in texts into structured data. into multiple subtypes of “fine-grained locations”
Information extraction is an effective way to [11]. Similarly, “fine-grained person”
populate the contents of a relational database. subcategories, like “politician” and “entertainer,”
Figure 2 illustrates how entity extraction can be appear in the work of [12]. The “person” type is
applied for text/data mining and visualization quite common and used at least once in an original
approaches. A typical information extraction way by [13], who combine it with other cues for
system has three to four major components [10]. extracting medication and disease names (e.g.,
The first component is tokenization or zoning “Parkinson disease”).
module. This module is used to split an input
document into its basic building blocks. The typical There are two known methods in NER. The first
building blocks are words sentences and method is based on knowledge-based approach and
paragraphs. The second component is a module for the second one is based on statistical or machine
performing morphological and lexical analysis. learning approach. Knowledge-based approach is
This module handles activities such as assigning also known as unsupervised learning approach,
post tags to the document’s various words. The while statistical and machine learning approach as a
third component is a module for syntactic analysis. supervised learning approach. In a supervised
This part of an IE system establishes the connection learning approach, NER process is learned
between the different parts of each sentence. automatically on large text corpora and then
supervised by human [14]. While in the
This is achieved either by doing full parsing or unsupervised learning approach, an existing lexical
shallow parsing, however in many IE works, a database such as WordNet is used [15]. The
shallow parsing approach is used. The forth combination of supervised and unsupervised
component is a module for domain analysis. The
learning approaches produces a semi-supervised while comprehension tests may require short
learning approach [16] [17]. answers phrases. To answer the question, the
system must be able to analyze the question first.
Although a significant number of NER research Questions can be distinguished by its type; factual,
work has been devoted to English language, some opinion or summary. If the question is in the
of other languages also have received attention by context of ongoing interaction, the answer must be
NER researchers. These include Turkish [18]), consulted with the online resource. Normally, the
Danish [19], Hindi [20], Polish [21], and Arabic answer is presented in some kinds of form [26].
language [22]. However, this task is quite
challenging in Arabic language because in this QA can be categorized into two types; according to
language, a proper name is not indicated by a its used methods. Firstly, shallow QA systems
capital letter (as in English language). Nevertheless, which use techniques like pattern matching in
some efforts have been shown for Arabic language. returning a final answer. Since such a method
For example, reference [23] has attempted to solve ignores the issue of semantic, thus, many relevant
the problem using a hybrid system built based on answers may be missed out or irrelevant answers
both statistical methods and predefined rules. may be retrieved. Secondly, deep QA systems; the
deep parsing technique of NLP is used [27].
Reference [24] presented an application that can
evaluate the performance of a number of the Arabic BASEBALL is the best known early question
root extraction methods. The implemented methods answering program [28]. The program answers
questions about baseball games played in the
in this system are selected according to a previous
American league over one season. Given a question
classification, where these methods are classified
into five categories: Light Stemmer, Arabic such as `Who did the Red Sox lose to on July 5?’ or
`How many games did the Yankees play in July?’
Stemming without a root dictionary, MT-based
or even on how many days in July did eight teams
Arabic Stemmer, N-gram based on similarity
play?, BASEBALL analysed the question, using
coefficient and N-gram based on dissimilarity
coefficient. The evaluation was conducted on the linguistic knowledge, into a canonical form which
was then used to generate a query against the
same terms in a corpus of two thousand words and
structured database containing the baseball data.
their roots.
One example of current working QA system is
Reference [25] presented a working Arabic START, the world’s first Web-based question
information extraction (IE) system that is used to answering system (http://start.csail.mit.edu/), has
analyze large volumes of news texts every day to been online and continuously operating since
extract the named entity (NE) types person, December, 1993. It has been developed by Boris
organization, location, date, and number, as well as Katz and his associates of the InfoLab Group at the
quotations (direct reported speech) by and about MIT Computer Science and Artificial Intelligence
people. In their work, they deployed multilingual Laboratory. START aims to supply users with "just
NER to cover Arabic where they presented what the right information,” instead of merely providing
Arabic language-specific resources had to be a list of hits. Currently, the system can answer
developed and what changes needed to be made to millions of English questions about places (e.g.,
the rule sets in order to be applicable to the Arabic cities, countries, lakes, coordinates, weather, maps,
language. demographics, political and economic systems),
movies (e.g., titles, actors, directors) and people
Question and Answering Systems (e.g., birth dates, biographies). A question such as
“Who is the ruler of Jordan?" can be given to the
Question and answering (QA) is a system system. The system will return an answer as the
that is able to automatically understand and build following” Ruler: King Abdullah II (1999)” and
answers to human user questions which are posed other relevant answers such as the current prime
in a natural language. The QA system is expected minister of Jordan. If a question such as `Why the
to allow users to ask questions in a "everyday sky is blue?’ is given, the system will give a long
language". Answers may be long or short, they may narrative answer. Reference [29] stated that START
be lists or narrative. They may vary with intended answers natural language questions by presenting
use and intended user [26]. For example, if a user components of text and multi-media information
wants a justification, this requires a longer answer, drawn from a set of information resources that are
hosted locally or accessed remotely through the of distilling the most important information from a
Internet. These resources contain structured, semi- source to produce a shorter version for a particular
structured and unstructured information. user or task”. Summarization is also considered as a
distilling process to obtain the most important
Although most of the published work reported on information from a source text to produce an
QA systems in English, some efforts in designing abridged version for a particular user/users and
and developing QA systems in Arabic language task/tasks. Therefore, the process of summarization
have been shown. AQAS has been considered as of any targeted text ranges from the interpretation
the earliest work for QA systems in Arabic of the source text, understanding and analyzing the
language [30]. The system was developed based on meaning being imparted to it; after that representing
knowledge based approach. It accepts Arabic the meaning of the source text which is based on
queries which follow pre-defined rules and matches certain areas of information, and at the end creating
these against frames in its knowledge base. The a summary of the source text that has been
architecture of the system is similar to early English understood and finalize all that needs to be
based QA systems. The system was applied in a represented.
closed domain environment; however, no
experimental results have been reported. The basic idea of summarization is to understand
the eventual meaning of the text which has been
Reference [31] reported on the design and presented in a short time and in the form of a
implementation of a question answering (QA) relatively shorter text. This shorter text, which is a
system called QARAB. QARAB takes natural subset of the original text, may not really convey all
language questions expressed in the Arabic the details of the actual text, but it certainly aims at
language and attempts to provide short answers. conveying the basic and actual idea which is trying
The systems primary source of knowledge is a to be conveyed in the text [34]. Research
collection of Arabic newspaper text extracted from community in this area have been putting effort to
Al-Raya, a newspaper published in Qatar. To apply automated text summarization in various
identify the answer, a keyword matching strategy applications such as document classification,
was adopted. The extracted keywords in the information retrieval, document extraction,
questions are matched to the candidate documents knowledge sharing, and so on. Text summarization
which are selected using information retrieval is really a complex task itself, since a wide variety
approach. QA system for Arabic language based on of techniques can be applied in order to condense
keyword matching is also reported in [32]. Using a content information, from pure statistical
keyword, simple structures extracted from both a approaches to those using closer analysis of text
question and candidate documents selected by the structure involving linguistic and heuristic methods
IR system were used in the process of identifying (anafora resolution, named entity recognition,
the answer. In order to perform this process, an lexical chains, etc,). In fact, many algorithms for
existing tagger is used to identify proper names and feature reduction, feature transformation, feature
other crucial lexical items and build lexical entries. weighting, etc. are directly related to this task, since
they already try to select a proper and limited set of
3.1.3 Automated Text Summarization items that can be used as storing the core content of
Automated text summarization is a process a given text.
of producing a readable, short, and meaningful
summary from a long text document. Automatic Most of the working summarization systems are
text summarization has drawn a considerable based on the extraction of a certain number of
interest since it provides a possible solution to a sentences found in the text which are considered to
critical information overload problem that people express most of the concepts present in the
are facing nowadays. An automated summarization document [35]. Sentence extraction techniques are
tool can be utilized by busy managers to scan usually statistical, linguistics, and heuristic
relevance information, researchers to have a quick methods, or a combination of all those techniques
glance on relevance research articles, students to in order to generate a final summary. The result is
have a quick understanding on subject matters, and not syntactically or content wise altered. In a
so on. Technically, summarization is a process of sentence extraction technique, a score is computed
deriving a shorter version of text from an original for each sentence based on features such as position
text, by selecting important contents. Text of sentence in the document and the frequency of
summarization was defined in [33] as "the process the word. Using these features, the most important
sentences in a document are extracted. A generated spam filtering, where email messages are classified
summary is basically a collection of extracted into the two categories of spam and non-spam,
sentences from the original text. Jusoh and her respectively. While text classification in the be-
research associates [36] have shown an effort to ginning was based mainly on heuristic methods
improve the sentence extraction technique by such as applying a set of rules based on expert
introducing the sentence refinement technique. The knowledge; nowadays the focus has turned to fully
summarization tool was developed and tested on automatic learning and even clustering methods.
texts which were written in English and Malay
language. Their experimental results indicated Some work on automated text summarization for
shorter versions of summaries are obtained without Arabic texts document can be found in the
losing its text context. published literatures [42] and [43]. For example,
the Arabic query-based text summarization system
Ultimately, the aim of summarization techniques is (AQBTSS) was reported in [42]. The system takes
to move one step forward; understanding and an Arabic document and a query (in Arabic) and
rearranging information of source texts to produce a attempts to provide a reasonable summary. In
readable and meaningful texts (sentence abstraction AQBTSS, each sentence is compared against the
technique) [37]. The challenging task is to user query to find relevant sentences. In this case,
understand the whole text and generate a summary the query has been used to select the documents. In
as a human does. Although this technique can 2009, El-Haj and associates presented a concept-
produce a better summary, however, this technique based summarizer system (ACBTSS) [43]. This
is very difficult to be implemented. Furthermore, in system takes a bag-of-words representing a certain
a manual abstraction process, an editor would have concept as the input the system instead of a user’s
to acknowledge six editing operations: reducing the query. Each sentence in a document is matched
sentences; combine them; transform them syntacti- against a set of keywords that represent a given
cally; paraphrasing its lexical; generalize and concept.
specify; and re-ordering the sentences [38]. Thus,
automated abstraction requires a computer system Challenges Issues in NLP
to have the knowledge of a human editor. Sentence Since decades, ambiguity has been a
abstraction requires lots of computing time. On the challenging issue for NLP researchers. In spite of
other hand, sentence extraction is easier to be some results on resolving ambiguity problems have
implemented as it does not require full been obtained, a number of important research
understanding of the texts context. problems have not been solved yet [44]. Ambiguity
is still a great challenge for computational linguists
Another possible technique is text categorization and computer scientists. The concept of ambiguity
(a.k.a. text classification) [39],[40] and [41]. is closely connected to semantic gap between the
Basically, text categorization is the task of user’s intentions and how she/he is able to convey,
assigning pre-defined categories to free-text since it can lead to more than one interpretation of
documents. Using this technique, a result of the the user’s input. Intentions are always a matter of
summarization algorithm is a list of key- interpretations. Ambiguity has been a critical issue
paragraphs, key-phrases or key-words that have for human computer interaction because of it
been considered to be the most relevant ones. pervasiveness in everyday life, yet its emergent
Although some methods are able to generate new nature challenges the role of design. Failure in
sentences from the content, usually it consists a giving a correct user’s interpretations may cause the
pure selection of textual fragments. It can provide user to mistrust the system and discontinue use.
conceptual views of document collections and has Several types of ambiguity have been identified.
important applications in the real world. For These include: structural, syntactical, form class,
example, news are typically organized by subject word sense and local ambiguity.
categories (topics) or geographical codes; academic
papers are often classified by technical domains and Structural Ambiguity
sub-domains; patient reports in health-care Structural ambiguity occurs when a
organizations are often indexed from multiple sentence can be analyzed as having more than one
aspects, using taxonomies of disease categories, syntactic structure or parse tree. For example, the
types of surgical procedures, insurance utterance “You can have peas and beans or
reimbursement codes and so on. Another carrots” can be analyzed in one of two ways, as
widespread application of text categorization is indicated by the bracketing [peas and beans] or
[peas and carrots]. Syntactical ambiguity is a PP attachment is the task of finding the governor
grammatical ambiguity of a whole sentence that for a given PP. In the given example above, the PP
occurs in sub-part-of a sentence. It is a grammatical with a spoon is governed by either the noun rice or
construct, and results from the difficulty of the verb eats. Determining the correct attachment
applying universal grammatical laws to a sentence site for PP is one of the major sources of ambiguity
structure. For example, a sentence “Salman hits the in natural language parsing and analysis.
boy with the stick”. This phrase is ambiguous, as to
whether a boy was hit with a stick, or whether a Early research on PP attachment focused on the
boy with a stick was struck by Salman. development of heuristics intended to model human
processing strategies, based on analysis of
Form class ambiguity arises when a given word can competing parse trees independent of lexical or
be analyzed as more than one part-of-speech. For discourse context. As the research communities
example, book, may be either a noun or a verb, grow up, many researchers have attempted to
plastic can be either an adjective or a noun, and so resolve PP attachment ambiguity in many different
on. Form class ambiguity necessarily gives rise to angles. A significant shift in NLP research on PP
structural ambiguity as well, as in the famous attachment was brought by the authors of [49] who
example “He saw her duck”. The words her and were the harbingers of statistical NLP and large-
duck are both form class ambiguous. Taking her as scale empirical evaluation. Researchers have been
possessive pronoun and duck as a noun, we get a trying to tackle the problem by a variety of
structure of [noun phrase, verb, noun, phrase]. smoothing methods and machine learning
However, taking her as a personal pronoun and algorithms including backed-off estimation [50],
duck as a verb, we get a structure of [noun phrase, instance-based learning 51], maximum entropy
verb, [noun phrase, verb]. learning [52], decision trees ([53], [54], neural
networks ([55], [56], boosting [57]) as well as
Structural disambiguation approach corpus [58].
The work on structural ambiguity reports
on preposition phrase (PP) attachment. The number Reference [59] also described a neural network
of published papers on lexical resolution is more based approach for resolving ambiguity in PP
than the number of published papers for structural attachment. To disambiguate the PP attachment, the
ambiguity. Prepositions are often among the most constituent namely verbs, noun, PP are associated
frequent words in a language. For example, based with semantic classes from WordNet. The neural
on the British National Corpus (BNC) [45], four out network method for PP attachment involves three
of the top-ten most-frequent words in English are phases, training, validation and testing. The
prepositions (of, to, in, and for). Despite their approach is classified into supervised learning
frequency, however, they are notoriously difficult to approaches. In their experiment, only one structure
master, even for humans [46]. For instance, less is used. A sentence does not go through a deep
than 10% of upper-level English as a Second parsing process. The work [58] proposed to resolve
Language (ESL) students can use and understand the PP attachment ambiguity based on a four-tuple
prepositions correctly [47]. composed of the head verb of the verb phrase, the
head noun of the noun phrase, and the preposition
Naturally, the number of PP contexts with and head noun in the prepositional phrase. A corpus
attachment ambiguity is theoretically unbounded. with known results, the Penn Treebank, is used for
The bulk of PP attachment research, however, has training and testing purposes.
focused exclusively on the case of a single PP
occurring immediately after an NP, which in turn is However, statistical approaches are not appropriate
immediately preceded by a verb. PP attachment or adequate in accounting for inferring
research has undergone a number of significant prepositional phrase attachments in cognitive
paradigm shifts over the course of the last three modeling systems, as human cognition is generally
decades, and been the target of interest of not a completely statistical process. Pure statistical
theoretical syntax, AI, psycholinguistics, statistical models for disambiguation tasks also suffer from
NLP, and statistical parsing [48]. Two large areas sparse-data problem. The hybrid method was
of research on the syntactic aspects of prepositions introduced in [60], [61], where in [60] a corpus-
are (a) PP attachment and (b) prepositions in based approach was integrated with knowledge-
multiword expressions. A sentence “Malik eats rice based techniques. In their work four head words
with a spoon” as an example of the PP attachment. were used; main verb (v), head noun (n1), the
preposition (p), and the head noun (n2), where it most appropriate senses with words in context”
was referred as quadruple (v, n1, p, n2). The clues [67].
include, syntactic cue, co-occurrance, syntactic
features and conceptual relationships between v and In [67] WSD was divided into two groups; lexical
n2 or between n1 and n2. They reported that the sample and all words WSD. In a lexical sample, a
results of their experiments are considered as good. system is required to disambiguate a restricted set
Reference [62] proposed a theoretical approach for of target words usually occurring one per sentence.
the detecting ambiguities connected with the In this type of systems, a number of instances are
meaning of the user’s input using a formal structure labeled manually (training set) and then applied to
for the multimodal input. The proposed approach is unlabeled instances (test set). This is also known as
also a hybrid approach which combines constraints a supervised system. In all words WSD, a system is
multiset grammar with linear logic. They claimed required to disambiguate all open-class words in a
that the hybrid approach provides an adaptive text. These include nouns, verbs, adjectives and
treatment of the ambiguities. adverbs. This task requires a wide coverage of
systems. Thus supervised systems can potentially
The importance of PP semantics has been discussed suffer from the problem of data sparseness, as it is
by many researchers. For example, [63] used unlikely that a training set of adequate size is
preposition semantics in a cooperative question available for a wide coverage. This is a point where
answering system in the context of cross-language the use of external knowledge is considered for
question answering (CLQA), and further later [64] WSD. This type of systems is classified into
successfully applied their preposition word sense unsupervised systems. Unsupervised systems based
disambiguation (WSD) method in a paraphrase their disambiguation decisions on knowledge
recognition task, namely, predicting that “Kim sources. The sources may belong to one of broad
covered the baby in blankets” and “Kim covered the classes: syntactic, semantic and pragmatic [68].
baby with blankets” have essentially the same Syntactic knowledge sources have to do with the
semantics. They proposed seven general senses of role of a word within the grammatical structures of
prepositions (e.g. PARTICIPANT, INSTRUMENT, sentences. Semantic knowledge relates the word to
and QUALITY), and annotated prepositions its properties. This was demonstrated by the work
occurring in 120 sentences for each of 10 of [69] where they have combined knowledge
prepositions. IE is one application where gathered from WordNet with results of an anaphora
prepositions are crucial to a system accuracy. As a resolution algorithm. Knowledge sources include
matter of fact, PP attachment plays an important corpora (a collection of text), machine readable
role in named entities and in IE patterns and in dictionaries and semantic network.
linking the elements in a text.
Knowledge-based approach
Word Sense Ambiguity
Lexical ambiguity or word sense
disambiguation (WSD) has been recognized as an The use of knowledge-based approach has
AI-hard problem. A break-through in this field been demonstrated in the early WSD work. For
would have a significant impact on many relevant example, [70] and [71] used manually encoded
Web-based applications, such as Web information semantic knowledge for WSD. Unfortunately, the
retrieval, improved access to Web services, IE, and manual creation of knowledge resources is an
so on [65]. WSD has obvious relationships to other expensive and time consuming effort, which must
fields such as lexical semantics, whose main be repeated every time the disambiguation scenario
endeavor is to define the relationships between changes. In recent years, existing lexical resources
“word” and “meaning” and “context” [66]). WSD such as machine-readable dictionaries (MRDs) like
can be viewed also as a classification task; word WordNet [72], [73], [67] and Oxford Dictionary of
senses are the classes, and an automatic English have been applied as an external source of
classification method is used to assign each knowledge in WSD work. According to [65] word
occurrence of a word to one or more classes based senses clearly fall under the category of objects that
on the evidence from the context and from external are better described through a set of structured
knowledge sources. WSD can be described as features. Thus they have applied structural pattern
“given a set of words (e.g., a sentence or a bag of recognition approach to disambiguate word senses.
words), a technique is applied which makes use of In their work, graph representations of word senses
one or more sources of knowledge to associate the are automatically generated from WordNet 2.7.
Others who used WordNet include researchers of
[74] [75] and [76]. Early approaches to WSD corpora and UMLS with the motivation that the
based on knowledge representation techniques, UMLS has been used in natural language
have been replaced in the past few years by more processing applications such as information
robust machine learning and statistical techniques. retrieval and information extraction systems. In
However, according to [65] the results of recent their work, machine-learning techniques have been
comparative evaluations of WSD systems show that applied to sense-tagged corpora, in which senses
machine learning and statistical techniques have (or concepts) of ambiguous terms have been most
inherent limitations. On the other hand, the manually annotated. Sense disambiguation
increasing availability of large-scale, rich lexical classifiers are then derived to determine senses (or
knowledge resources seems to provide new concepts) of those ambiguous terms automatically.
challenges to knowledge-based approaches [67]. However, they conclude that manual annotation of
a corpus is an expensive task.
Machine learning approach
The work presented in [77] has been Research of [82] proposed a method for lexical
considered as an early application of machine ambiguity resolution using corpus and concept
learning to the WSD problem. Several information. Since the extracted knowledge is
disambiguation cues, such as first noun to the stored in words themselves, these methods require a
left/right and second word to the left/right were large amount of space with a low recall rate. On the
extracted from parallel text. The senses are defined contrary, they resolve word sense ambiguity by
by determining the differences between them. This using concept co-occurrence information extracted
technique was also applied for machine translation. from an automatically sense-tagged corpus. The
On the other hand, [78] used the flip-flop algorithm tested accuracy of their method exceeds 82.4% for
to decide which of the important cues for each word nominal words, and 83% for verbal words.
by using mutual scores between words. Syntactic
relations between subject-verb, verb-object and Although WSD have not been applied to real task
adjective-noun have been used by [79] to determine applications widely, a few researchers have taken
the cues. an effort to do so. For example, reference [83] have
extended the previous work by mining biological
According to [80], most of the previous corpus- named entity tagging (BNET) that identifies names
based approaches to the resolution of word-sense mentioned in text and normalizes them with entries
ambiguity are based on lexical information from the in biological databases. They concluded that that
context of the word to be disambiguated suffer from names for genes/proteins are highly ambiguous and
the problem of data sparseness. To address this there are usually multiple names for the same gene
problem, they proposed a disambiguation method or protein. Reference [84] investigated a particular
using co-occurring concept codes (CCCs). The use technique for resolving ambiguity that is motivated
of concept-code features and concept-code by task-level ambiguity. In their study, they
generalization effectively alleviate the data explored a technique to find commonly occurring
sparseness problem and also reduce the number of patterns of part-of-speech in a query and allow the
features to a practical size without any loss in patterns to be transformed into clarification
system performance. They claimed that the questions. The patterns are centered on a single
effectiveness of the CCC features and the concept- query word and incorporate a small number of
code generalization by experimental evaluations. words on either side.
The proposed disambiguation method was applied
to a Korean-to-Japanese MT system that 4. THE FUTURE OF NLP
experimented with various machine-learning
techniques. In a lexical sample evaluation, their Despite the lack of high-performing methods
CCC-based method achieved a precision of had preventing an extensive use of NLP techniques
82.00%, with an 11.83% improvement over the in many areas of information technology, such as
baseline. Also, it achieved a precision of 83.51% in information retrieval, natural language interfaces,
an experiment on real text, which shows that their query processing, advanced Web search, and many
proposed method is very useful for practical MT more real applications, NLP based applications are
systems. emerging technologies for business.

The work presented in [81] demonstrated an effort It is no doubt that NLP is an enabler for deploying
to resolve ambiguous terms using sense-tagged natural, intelligent, and intuitive applications for
everyday use. It is transforming the way how technologies which are based on NLP are briefly
human interact with computers. Thus, resolving the highlighted.
complexity issues in a human language is indeed
critical, vital, and urgent. REFRENCES:
[1] McCallum, A. (2005). Information extraction:
Applications such as chatbot, smart search, distilling structured data from unstructured text.
recommender, customer service, personal assistant, Queue, 3 , 48–57.
multi lingual automated translation machine, [2] Sekimizu, T., Park, H., & Tsuji, J. (1998).
question answering, caption generation are Identifying the interactions between genes and
expected to be able to capitalize NLP techniques gene products based on frequently seen verbs in
for human-like understanding of speech and texts. medline abstract. Tokyo Japan: Universal
Deeper applications such as extracting insights and Academy Press.
analysis from a vast amount unindexed and
[3] Chu, C.-T., Sung, Y.-H., Yuan, Z., & Jurafsky,
unstructured data, mining texts, images, audios and
D. (2006). Detection of word fragments in
videos or reading, filtering, analyzing, extracting,
mandarin telephone conversation. In
and visualizing pieces of knowledge from text
International Conference on Spoken Language
documents such as emails, short messages, reviews,
Processing. URL pubs/fragment-icslp-06.pdf
and so on, are seen as critical technologies of NLP
in the future. [4] Ramage, D., Rosen, E., Chuang, J., Manning, C.
D., & McFarland, D. A. (2009). Topic modeling
When machines are intelligent enough to for the social sciences. In Workshop on
understand and communicate in a human language, Applications for Topic Models: Text and
human users are able to be more effective and Beyond (NIPS 2009 ). Whistler, Canada.
efficient in accessing, analyzing, and leveraging [5] Jurafsky, D., Ranganath, R., & McFarland, D.
huge amount of data. NLP market is growing. (2009). Extracting social meaning: identifying
According to a 2017 Tractica report [86], NLP interactional style in spoken conversation. In
market is estimated to be around 22.3 billion USD Proceedings of Human Language Technologies:
by 2025. This estimation has included the total NLP The 2009 Annual Conference of the North
software, hardware and services. Furthermore, NLP American Chapter of the Association for
solutions that leveraging AI will see a market Computational Linguistics (NAACL ’09), (pp.
growth from 136 million USD in 2016 to 5.4 billion 638–646). Morristown, NJ, USA: Association
USD by 2025. for Computational Linguistics.
[6] Grenager, T., Klein, D., & Manning, C. D.
5. CONCLUSION (2005). Unsupervised learning of field
segmentation models for information extraction.
This paper has successfully present the most In Proceedings of the 43rd Annual Meeting on
prominent applications of NLP. These include Association for Computational Linguistics, (pp.
information extraction, question answering 371–378).
systems, and automated text summarizations. [7] Jurafsky, D., & Martin, J. H. (2009). Speech and
Because of the mechanism (the citation numbers) Language Processing: An introduction to
used in selecting papers to be reviewed, a number Natural language Processing, Computational
of current literatures which discussed the selected Linguistics and Speech Recognition. United
topic might be left out unintentionally. In overall, States of America: Prentice Hall.
this paper has given a depth overview of main [8]Allen, J. (1988). Natural Language
applications of NLP. NLP has been also considered Understanding, United States of America: The
as one of AI hard problems. The complexity of Ben-jamin/Cummings Publishing Company.
natural language processing is caused by the
[9] Karat, C., Vergo, J., & Nahamoo, D. (2003).
ambiguity problems which always occur in a
Conversational interface technologies. In J. A.
human language. Although the ambiguity problem
Jacko, & A. Sears (Eds.). The Human-
may occur in all levels of a natural language, the
Computer Interaction Handbook, (pp. 169–186).
most common problems always occur at lexical and
Lawrence Erlbaum Associates.
structural levels. The paper also addresses,
discusses, and distinguishes between approaches in [10] Feldman, R., & Sanger, J. (2007). The text
resolving the ambiguity problems. The future mining Handbook: Advanced Approaches in
 Analyzing Unstructured Data. United State of [22] Huang, F. (2005). Multilingual Named Entity
America: Cambridge University Press. Extraction and Translation from Text and
[11] Lee, S., & Lee, G. (2005). Heuristic methods for Speech. Ph.D. thesis, Carnegie Mellon
reducing errors of geographic named entities University.
learned by bootstrapping. In Proceeding of the [23] Abuleil, S. (2006). Hybrid system for extracting
International Joint Conference on Natural and classifying arabic proper names. In
Language Processing. Proceedings of the 5th WSEAS International
[12] Fleischman, M., & Hovy, E. (2002). Fine Conference on Artificial Intelligence,
grained classification of named entities. In Knowledge Engineering and Data Bases, (pp.
Proceeding of the 19th International Conference 205–210). Stevens Point, Wisconsin, USA:
on Computational Linguistics (COLING). World Scientific and Engineering Academy and
[13] Bodenreider, O., & Zweigenbaum, P. (2000). Society (WSEAS).
Identifying proper names in parallel medical [24] AlHajjar, A., Hajjar, M., & Khaldoun, Z.
terminologies. Stud Health Technol Inform, 77 , (2010). A system for evaluation of arabic root
443–447. extraction methods. In Proceedings of the 2010
[14] McCallum, A., & Li, W. (2003). Early results Fifth International Conference on Internet and
for named entity recognition with conditional Web Applications and Services, ICIW ‟10, (pp.
random fields, feature induction and Web- 506–512). Washington, DC, USA: IEEE
enhanced lexicons. In Proceedings of the Computer Society.
Conference on Computational Natural [25] Zaghouani, W. (2012). Renar: A rule-based
Language Learning . arabic named entity recognition system. 11 (1),
[15] Alfonseca, E., & Manandhar, S. (2002). An 2:1–2:13.
unsupervised method for general named entity [26] Hirschman, L., & Gaizauskas, R. (2001).
recognition and automated concept discovery. Natural language question answering:the view
In Proceedings of the 1st International from here. Natural Language Engineering, 7,
Conference on General WordNet, (pp. 466– 275–300.
471). [27] Al-Harbi, O., Jusoh, S., & Norwawi, N. M.
[16] Chang, C. H., & Kup, S.-C. (2004). A semi- (2011). Lexical disambiguation in natural
supervised approach of web data extraction with language questions-nlqs). Journal of Computer
visual support. Intelligent System, 19 (6), 56–64 Science Issues, 8, 143 International –150.
[17] Nadeau, D. (2007). Semi-Supervised Named [28] Green, B., Wolf, A., Chomsky, C., & Laughery,
Entity Recognition:Learning to Recognize 100 K. (1961). Baseball: An automatic question
Entity Types with Little Supervision. Ph.D. answerer. In Proceedings Western Joint
thesis, University of Ottawa. Computer Conference, vol. 19, (pp. 219–224).
[18] Cucerzan, S., & Yarowsky, D. (1999). [29] Katz, B., Borchardt, G., & Felshin, S. (2006).
Language independent named entity recognition Natural language annotations for question
combining morphological and contextual answering. In Proceedings of the 19th
evidence. In Proceedings of the Joint Sigdat International FLAIRS Conference (FLAIRS
Conference on Empirical Methods in Natural 2006).
Language Processing and Very Large Corpora. [30] Mohammed F. A, Khaled Nasser, & Harb H.M.
[19] Bick, E. (2004). A named entity recognizer for (1993). A knowledge based Arabic question
Danish. In Proceedings of the Conference on answering system (AQAS). SIGART Bull. 4, 4
Language Resources and Evaluation. (October 1993), 21-30.
[20] May, J., Brunstein, A., Natarajan, P., & [31] Hammo, B., Abu-Salem, H., & Lytinen, S.
Weischedel, R. (2003). Surprise! what‟s in a (2002). Qarab: a question answering system to
cebuano or Hindi name? ACM Transactions on support the arabic language. In Proceedings of
Asian Language Information Processing the ACL-02 workshop on Computational
(TALIP), 2 (3), 169–180 approaches to semitic languages, SEMITIC ‟02,
[21] Piskorski, J. (2004). Named-entity recognition (pp. 1–11). Stroudsburg, PA, USA: Association
for Polish with SProUT. In L. Bolc, Z. for Computational Linguistics.
Michalewicz, & T. Nishida (Eds.) Lecture [32] Kanaan, G., Hammouri, A., Al-Shalabi, R., &
Notes in Computer Science, vol. 3490, (pp. Swalha, M. (2009). A new question answering
122–133). system for the arabic language. American
Journal of Applied Sciences, 6, 797–805.
[33] Mani, I., & Benjamin, J. (2002). Review of [44] Alfawareh, H.M. & Jusoh, S. (2011). Resolving
automatic summarization. Journal of ambiguous entity through context knowledge
Computational Linguistics, 28, 221–223. and fuzzy approach. International Journal on
[34] Mani, I. (1999). Advances in Automatic Text Computer Science and Engineering (IJCSE), 3
Summarization. Cambridge, MA: MIT Press. (1), 410 – 422.
[35] Loo, P., & Tan, C. (2002). Word and sentence [45] Burnard, L. (2000). Reference Guide for the
extraction using irregular pyramid. In British National Corpus. Oxford, UK: Oxford
Proceedings of the 5th International Workshop University Computing Services.
on Document Analysis Systems V (DAS ‟02), [46] Chodorow, M., Tetreault, J., & N.Han (2007).
(pp. 307–318). Heidelberg: Springer-Verlag, Detection of grammatical errors involving
London, UK. prepositions. In Proceedings of the 4th ACL-
[36] Jusoh, S., Masoud, A. M., & Alfawareh, H. M. SIGSEM Workshop on Prepositions, (pp. 25–
(2011). Automated text summarization: 30).
Sentence refinement approach. In V. Snasel, J. [47] Lindstromberg, S. (2001). Preposition entries in
Platos, & E. El-Qawasmeh (Eds.) Digital UK monolingual learners dictionaries: Problems
Information Processing and Communications, and possible solutions. Applied Linguistics, 22
vol. 189 of Communications in Computer and (1), 79–103.
Information Science, (pp. 207–218). Springer [48] Baldwin, T., Kordoni, V., & Villavicencio, A.
Berlin Heidelberg. (2009). Prepositions in applications: A survey
[37] Chan, S. (2006). Beyond keyword and cue- and introduction to the special issue.
phrase matching: a sentence-based abstraction Computational Linguistic, 35 (2), 119–149.
technique for information extraction. Decision [49] Hindle, D., & Rooth, M. (1993). Structural
Support System, 42, 759–77. Chang, C. H., & ambiguity and lexical relations. Computational
Kup, S.-C. (2004). A semi-supervised approach Linguistics., 19 (1), 103–120.
of web data extraction with visual support. [50] Collins, M., & Brooks, J. (1995). Prepositional
Intelligent System, 19 (6), 56–64. phrase attachment through a backed-off model.
[38] Jeek, K., & Steinberger, J. (2008). Automatic In Proceedings of the 3rd Annual Workshop on
text summarization: The state of the art and new Very Large Corpora, (pp. 27–38).
challenges. In Proceedings of the Znalosti 2008, [51] Zavrel, J., Daelemans, D., & Veenstra, J.
(pp. 1–12). (1997). Resolving PP attachment ambiguities
[39] Devasena, C. L., & Hemalatha, M. (2012, with memory-based learning. In Proceedings of
March). Automatic text categorization and the Conference on Computational Natural
summarization using rule reduction. Language Learning (CoNLL-97), (pp. 136–
In Advances in Engineering, Science and 144).
Management (ICAESM), 2012 International [52] Ratnaparkhi, A., Reynar, J., & Roukos, S.
Conference on (pp. 594-598). IEEE. (1994). A maximum entropy model for
[40] Sebastiani, F. (2002). Machine learning in preposi¬tional phrase attachment. In
automated text categorization. ACM computing Proceedings of the workshop on Human
surveys (CSUR), 34(1), 1-47. Language Technol¬ogy, (pp. 250–255).
[41] Yang, Y. (1999). An evaluation of statistical [53] Merlo, P., Crocker, M. W., & Berthouzoz, C.
approaches to text categorization. Information (1997). Attaching multiple prepositional
retrieval, 1(1), 69-90. phrases: Generalized backed-off estimation. In
[42] El-Haj, M., & Hammo, B. (2008). Evaluation of Proceedings of the 2nd Conference on
query-based arabic text summarization system. Empirical Methods in Natural Language
In Proceeding of the IEEE International Processing (EMNLP-97), (pp. 149–155).
Conference on Natural Language Processing [54] Alam, Y. S. (2004). Decision trees for sense
and Knowledge Engineering, NLP-KE08, (p. disambiguation of prepositions: Case of over. In
17). IEEE Computer Society. Proceedings of the Workshop on Computational
[43] El-Haj, M., Kruschwitz, U. and Fox, C., 2009, Lexical Semantics, (pp. 52–59).
November. Experimenting with Automatic Text [55] Sopena, J. M., Lloberas, A., & Moliner, J. L.
Summarisation for Arabic. In LTC (pp. 490- (1998). A connectionist approach to
499). prepo¬sitional phrase attachment for real world
texts. In Proceedings of the 36th Annual
 Meeting of the ACL and 17th International Conference on Intelligent Text Processing and
Conference on Computational Linguistics Computational Linguistics (CICLing-2006),
(COLING/ACL-98), (pp. 1233–1237). (pp. 196–207).
[56] Alegre, M. A., Sopena, J. M., & Lloberas, A. [65] Navigli, R., & Velardi, P. (2005). Structural
(1999). PP-attachment: A committee machine semantic interconnections: a knowledge-based
approach. In Proceedings of the Joint SIGDAT approach to word sense disambiguation. IEEE
Conference on Empirical Methods in Natural Transactions on Pattern Analysis an
Language Processing and Very Large Corpora [66] Agirre, E., & Edmonds, P. (2007). Introduction.
(EMNLP/VLC-99), (pp. 231–238). In E. Agirre, & P. Edmonds (Eds.) Word Sense
[57] Abney, S., Schapire, R. E., & Singer, Y. (1999). Disambiguation: Algorithms and Applications,
Boosting applied to tagging and pp attachment. (pp. 1–28). New York: Springer Verlag.
In Proceedings of the Joint SIGDAT [67] Navigli, R. (2009). Word sense disambiguation:
Conference on Empirical Methods in Natural a survey. ACM Computing Surveys, 41 (2), 1–
Language Processing and Very Large Corpora 69.
(EMNLP/VLC-99), (pp. 38–45). [68] Agirre, E., & Stevenson, M. (2007). Knowledge
[58] Nadh, K., & Christian, H. (2009). Prepositional sources for WSD, (pp. 217–251). New York:
phrase attachment ambiguity resolution using Springer Verlag
semantic hierarchies. In Proceedings of the [69] McCarthy, D., Carroll, J., & Preiss, J. (2001).
Ninth IASTED International Conference on Disambiguating noun and verb senses using
Artificial Intelligence and Applications, (pp. automatically acquired selectional preferences.
73–80). In Proceedings of the SENSEVAL-2 Workshop
[59] Srinivas, M., & Bhattacharyya, P. (2006). at the European Chapter ACL, (pp. 119–122).
Prepositional phrase attachment through Toulouse, France.
seman¬tic association using connectionist [70] Schank, R., & Abelson, R. (1977). Scripts,
approach. In Proceedings of the Third Plans, Goals, and Understanding. Hillsdale, N.J:
International WordNet Conference Lawrence Erlbaum.
(GWC2006), (pp. 273–277).
[71] Wilks, Y. (1978). A preferential pattern-seeking
[60] Wu, H., & Furugori, T. (1996). Prepositional semantics for natural language inference.
phrase attachment through a hybrid Artificial Intelligence, 6 , 53–74.
disambiguation model. In Proceedings of the
[72]Resnik, P. (1995). Using information content to
16th conference on Computational linguistics,
evaluate semantic similarity in a taxonomy. In
(pp 1070-1073). Morristown, NJ, USA:
Proceedings of the International Joint
Association for Computational Linguistics.
Conference Artificial Intelligence (IJCAI), (pp.
[61] Hartrumpf, S. (1999). Hybrid disambiguation of 448–453).
prepositional phrase attachment and in-
[73] Mihalcea, R., & Moldovan, D. (2001). A highly
terpretation. In Proceedings of the Joint
accurate bootstrapping algorithm for word sense
SIGDAT Conference on Empirical Methods in
disambiguation. International Journal of
Natural Language Processing and Very Large
Artificial Intelligence Tools, 10 (1-2), 5– 21.
Corpora (EMNLP/VLC-99), (pp. 111–120).
[74] Jiang, J. J., & Conrath, D. W. (1997). Semantic
[62] Chiara, C. M., Fernando, F., & Patrizia, G.
similarity based on corpus statistics and lexical
(2008). Ambiguity detection in multimodal
taxonomy. In Proceedings of the 10th
systems. In Proceedings of the Working
International Conference on Research in
Conference on Advanced visual interfaces, (pp.
Computational Linguistics.
331–334). New York, NY, USA: ACM.
[75] Agirre, E., & Martinez, D. (2000). Exploring
[63] Benamara, F. (2005). Reasoning with
automatic word sense disambiguation with
prepositions within a cooperative question-
decision lists and the web. In Proceedings of the
answering framework. In Proceedings of the
Semantic Annotation And Intelligent
Second ACL-SIGSEM Workshop on the
Annotation workshop organized by COLING
Linguistic Dimensions of Prepositions and their
Luxembourg 2000 , (pp. 11–19).
Use in Computational Linguistics Formalisms
and Applications, (pp. 145–152). [76] Banerjee, S., & Pedersen, T. (2003). Extended
gloss overlaps as a measure of semantic relat-
[64] Boonthum, C., Toida, S., & Levinstein, I.
edness. In Proceedings of the 18th international
(2006). Preposition senses: Generalized
disambiguation model. In Proceedings of the
 joint conference on Artificial intelligence,
(pp.805–810).
[77] Brown, P., Stephen, E., Pietra, D., Vincent, J.,
Pietra, D., & Mercer, R. L. (1991). Word sense
disambiguation using statistical methods. In
Proceedings of the 29th Annual Meeting for
Computational Linguistics, (pp. 264–270).
[78] Nadas, A., Nahamoo, D., Picheny, M., &
Powell, J. (1991). An iterative approximation of
the most informative split in the construction of
decision trees. In Proceedings of the IEEE
International Conference on Acoustics, Speech
and Signal Processing, (pp. 565– 568). Toronto.
[79] Yarowsky, D. (1996). Homograph
disambiguation in text-to-speech synthesi. In J.
Hirschberg, R. Sproat, & J. van Santen (Eds.)
Progress in Speech Synthesis, (pp. 159–175).
New York: Springer Verlag. [77]
[80] Youjin, C., & Jong-Hyeok, L. (2005). Practical
word-sense disambiguation using c-occurring
concept codes. Machine Translation, 19 (1), 59–
82.
[81]Liu, H., Hu, Z., Torii, M., Wu, C., Friedman, C.
(2006). Quantitative assessment of dictionary-
based protein named entity tagging. Journal of
the American Medical Informatics Associations
(JAMIA), 13, 497–507.
[82] Liu, H., Johnson, S. B., & Friedman, C. (2002).
Automatic resolution of ambiguous terms based
on machine learning and conceptual relations in
the UMLS. Journal of the American Medical
Informatics Associations (JAMIA), 9, 621–636.
[83] Liu, H., Hu, Z., Torii, M., Wu, C., & Friedman,
C. (2006). Quantitative assessment of
dictionary-based protein named entity tagging.
Journal of the American Medical Informatics
Associations (JAMIA), 13, 497–507.
[84] James, A., & Hema, R. (2002). Using part-of-
speech patterns to reduce query ambiguity. In
Proceedings of the 25th annual international
ACM SIGIR conference on Research and
development in information retrieval, (pp. 307–
314). New York, NY, USA: ACM.
[85]Tractica, Natural Language Processing Market
to Reach $22.3 Billion by 2025, August 21,
2017, Retrieved from:
https://www.tractica.com/newsroom/press-
releases/natural-language-processing-market-to-
reach-22-3-billion-by-2025/