ARABIC LANGUAGE CHALLENGES IN TEXT BASED CONVERSATIONAL AGENTS COMPARED TO THE ENGLISH LANGUAGE

International Journal of Computer Science & Information Technology (IJCSIT) Vol 7, No 3, June 2015
DOI:10.5121/ijcsit.2015.7301 1
ARABIC LANGUAGE CHALLENGES IN TEXT BASED
CONVERSATIONAL AGENTS COMPARED TO THE
ENGLISH LANGUAGE
Mohammad Hijjawi1
and Yousef Elsheikh 2
1
Department of Computer Science, Applied Science University, Amman, Jordan
2
Department of Computer Information Systems, Applied Science University, Amman,
Jordan
ABSTRACT
This paper is not to compare between the Arabic language and the English language as natural languages.
Instead, it focuses on the comparison among them in terms of their challenges in building text based
Conversational Agents (CAs). A CA is an intelligent computer program that used to handle conversations
among the user and the machine. Nowadays, CAs can play an important role in many aspects as this work
figured. In this paper, different approaches that can be used to build a CA will be differentiated. In each
approach, the comparison aspects among the Arabic and English languages will be debated with the
respect to the Arabic language.
KEYWORDS
Conversational Agents, Chatterbots, Arabic Language and English Language.
1. INTRODUCTION
A Conversational Agent is an intelligent computer program has been developed to handle
utterances/conversations among the user and the machine. CAs has been started with the idea of
Alan Turing introduced which is Turing test [1]. Turing test or Turing game modelled to test the
human’s ability of differentiating among the second conversation party character. The human will
chat with the second party which might be another human or a machine via textual conversations.
The game consider the first human fail if he/she failed to judge that the second party in the
conversation was a human or a machine. From Turing’s time, many of CAs has been proposed
and developed until these days such as ELIZA [2], ALICE [3], ADAM [4], Museum Guide[5],
InfoChat[6], InCA[7], Abdullah [8]andArabChat[9]. Most of the previous CAs was for English
language. A manual listing that the authors built for the current CAs appeared that more than 97%
of current CAs is not for the Arabic language and most of them was for the English. To best of
our knowledge, this is due to three main factors; firstly, Arabs has adopt the Internet in general in
1990s which might be considered late comparing to the English based countries. This late
affected the research and adopting the technology in general. Secondly, the English language has
the first rank in terms of number of speakers worldwide [10] which lead the English based
countries and the worldwide companies to focus building their CAs for the English language.
Thirdly, the difficulties and challenges that the Arabic language facing.
This paper discusses the challenges that the Arabic based CAs face comparing to the English
based CAs.

2
The Arabic language is a Semitic language that originated in the period between the seventh
century BC and the third century AD in the Middle East region [11]. Arabic, like other languages,
affects and is affected by other languages in terms of accepting new vocabulary in its dictionary
especially from the technology field. Arabic spread significantly when Islam spread to Asia,
Africa and Europe. This consequently increased the number of Arabic speakers in those regions,
especially because it is the spiritual language of Islam [11].Presently, different types of Arabic
language are used in daily conversations between people. These are known as Classical, Modern,
and Colloquial Arabic [11]. Arab speakers usually use these different types of Arabic depending
on the nature of conversation and sometimes in the same conversation which caused a diglossia.
The Diglossia is a phenomenon whereby two or more kinds of the same language exists in the
same speech community [12].
Classical Arabic, as used in the eras before Islam and in the Quran, is more complex in its
grammar and vocabulary than modern Arabic. Classical Arabic has a large number of diacritics
that facilitate word pronunciation and detection in their grammatical cases (for instance, noun or
verb). In the Modern Arabic, these diacritics were omitted to accelerate the reading and writing
processes. Modern Arabic is formally recognised as the official language of the Arab countries. It
is used in everyday language, in the media, education, and literature. Most of computational
Arabic-based research use Modern Arabic [13]. The third type, known as Colloquial Arabic, is
less sophisticated than Modern Arabic in its grammar and vocabulary. However, due to its
simplicity, most people use it in their everyday spoken conversations and in informally written
letters [14].
Arab people have weaknesses in using Modern Arabic which make them generate incorrect
sentences in terms of grammar. In addition, they might mix between Modern and Colloquial
Arabic, which increases the challenge for a CA to understand or recognise the user’s utterance.
Moreover, different Arab countries have different Colloquial Arabic or “dialects”, but most these
countries can understand each other. This will might increase the challenge for an Arabic CA to
understand or recognise user utterances from various Arab countries when they used their
Colloquial Arabic. In contrast, this problem will be smaller in the English language as it
nowadays has mainly two types which are the standard English and colloquial English. However,
the two types are closed to each other in grammar and the sentence’s structure. In addition, the
standard English is the same in all countries that speak it such as America, United Kingdom and
Canada. Although, the colloquial English is differ from country to country but this difference is in
the accent (the spoken language) not in the written text language.
Arabic has 28 letters, each of which has many written forms depending on their position in the
word (initial, middle, or end) such as the letter ‘‫’ب‬ ‘b’ which is in the independent shape has three
forms in Arabic which are ‘‫ـ‬ ’ when it is written in first of a word, and ‘‫ـ‬ ‫’ـ‬ when it is written in the
middle and ‘ ‫’ـ‬ when it is written at the end of a word. In contrast, the letter ‘b’ has the same
form wherever written in a word. These different forms are due to the Arabic writing style
(concatenative and cursive). Arabic differs from English in terms of the direction of writing, it
iswritten from right to left [16]. Where the English written from the left to right. Moreover,
Arabic does not have capital or small letters, and does not support capitalisation features. These
missed features in Arabic will increase the processing challenges as will discuss in this paper.
Generally, linguistic textual CAs has three main approaches to build which are; Natural Language
Processing based approach, sentence similarity measures based approach and Pattern Matching
approach. These approaches can be used to build the Arabic based and English based CAs.
However, all of them has different research history for Arabic and English but the English
language research amount has the majority of this research history. The next section describes the
variances among these approaches for the Arabic and English.

3
2. CONVERSATIONAL AGENT APPROACHES CHALLENGES FOR THE ARABIC
AND ENGLISH
As mentioned above, three main approaches can be adopted to develop CAs, namely, Pattern
Matching (PM), Natural Language Processing (NLP), and Sentence Similarity Measures
(SSM).Researchers should take care when they are deciding which approach they are going to
choose in order to build their CA in terms of the compatibility between the used approach and the
targeted natural language. In general, each approach has advantages and disadvantages. These
advantages and disadvantages might be increased or decreased depending on the used approach
with a specific natural language. In other words, an approach that could be useful when used with
one natural language might not be useful enough when used with another natural language for
number of reasons as the following subsections investigate.
2.1 Natural Language Processing based CA
NLP is defined in computational linguistics as “the computational processing of textual materials
in natural human languages”[17]. NLP might also be defined as studying the constructing and
meanings of a natural language through applying its rules to process information enclosed within
its sentences [18]. The general aim of NLP is to develop computational techniques, which will
analyse huge number of spoken or written texts in the manner of human carrying out the same
task [17]. NLP emerged in the 1950s with machine translation, and it had further focusing by
research in artificial intelligence field. More concentration has been placed on expert systems or
intelligent programs to simulate human behaviour and human’s knowledge in order to make
inferences and interpretations and then reach conclusions like human.
NLP seems to be the most effective way to build CAs especially when its techniques try to deal
with utterances semantically, which means understanding the utterance’s subject and its content.
Understanding a sentence needs to understand the sentence’s structure, its components (words)
and the relations among those components.
Arabic and English has a linguistic morphology features. According to [13], a “Morphology is the
study of internal word structure”. It is usually focuses on two fundamental issues: firstly,
derivational morphology, which concerns how words are formed; and secondly, inflectional
morphology, which concerns how words interact with the syntax [11]. However, derivational
morphology governs the principles of a word’s transformation from its root. For instance, from
the root “ ”, many different words might be generated such as “ ” “he is writing”, “ ‫”ﺗ‬
“she is writing”, “‫ب‬ ” “written or a letter”, and “ ” “library”. This richness of generating
words with different semantic meanings presents a challenge to any kind of Arabic based
computational processing.Morphological features work upon integrated dependencies among
several linguistic factors, such as vowels, affixes and root-based systems, as the rest of this
section will describe.
Affixes and root base system
Semitic languages such as Arabic have a very rich derivational and inflectional features that are
based on a root to generate the language’s words. A root is the initial form of a word that cannot
be analysed further [15]. For instance, from the root “ ”, many different words might be
generated such as “ ” “he is writing”, “ ‫”ﺗ‬ “she is writing”, “‫ب‬ ” “written or a letter”, and
“ ” “library”.
The Arabic language is based on a root system to generate its words. Arabic has more than 10,000
roots [15]. According to Al-Fedaghi and Al-Anzi[19], “85% of words are derived from tri-literal
roots”; the rest are derived from quad-literal, pent-literal and hex-literal roots [15].

4
Affixes such as prefixes, infixes, and suffixes can be added before, inside, or after a root
respectively, to generate more meaningful words. Affixes in Arabic such as “‫,”ا‬ “‫,”ال‬ “‫,”وال‬ “‫,”ب‬
“‫ل‬ ”, “‫,”ن‬ “‫”ون‬ and “ ‫”وا‬ can be generated from the combination of letters of the word “ ”
“You asked her” [13].
Arabic derivational and inflectional morphology features are based on affixes. For instance, the
derivational morphology feature can generate words such as “‫رس‬ , ‫روس‬ , ‫”دارس‬ from the root
“‫”درس‬ by mapping the root with Arabic word-based patterns. Inflectional morphology, which are
generated by adding prefixes or suffixes to a root, may reflect a tense (past/present), a gender
(masculine, feminine) and/or a number (singular, dual, plural). For instance, from the root “‫”درس‬
many words might be generated, such as “‫ن‬ ‫ر‬ “teachers”‫درس‬ “I taught”, ‫ن‬ ‫ر‬ ”two teachers
or two schools”, ‫ر‬ ”a school” ”.
Arabic grammar has a number of powerful structures to govern word transformation, for instance
words patterns. Words patterns are words formed based on affixation and vocalisation processes
on the Arabic roots. They can be classified into noun patterns and verb patterns [15] and they
formed based upon the essential pattern “ !”. For instance, the word “ "” “Worker” belongs to
pattern ( " !), and “ ” “He works” belongs to pattern ( # ). The generated words from the
same root might not be related to each other semantically, which increases the ambiguity in
Arabic semantic processing [20]. For instance, the root “$ %” can generate the word “$ َ%” which
means “month”, and “‫ر‬ 'َ ” which means “famous”. Table 1 shows an example of how
derivational and inflectional morphology features can reform the root (“ (‫.)”ط‬ In Table 1, adding
the infix “‫”ا‬ to the mentioned root, a new word will be generated which is “ ‫”ط‬ “a male
student”, while adding the prefix “‫”ي‬ to the root, it will generate the word “ (+ ” “He requests”.
These rich morphological features will increase the challenge for a CA to determine the
grammatical case for a user’s utterance words. This might leads to an inability to understand or
recognise the user utterance for CAs based on NLP.
Table 1: Examples of how derivational and inflectional morphology can reform a root.
Arabic word Prefix Infix Suffix Stem Root English Translation
, + ‫ا‬ ‫ال‬ ‫ا‬ , ‫ط‬ (‫ط‬ Students (dual, masculine)
, + ‫ا‬ ‫ال‬ ‫ا‬ , ‫ﺗ‬ ‫ط‬ (‫ط‬ Students (dual, feminine)
‫ن‬ + ‫ا‬ ‫ال‬ ‫ا‬ ‫ان‬ ‫ط‬ (‫ط‬ Students (dual, masculine)
‫ن‬ + ‫ا‬ ‫ال‬ ‫ا‬ ‫ن‬ ‫ﺗ‬ ‫ط‬ (‫ط‬ Students (dual, feminine )
‫+-ب‬ ‫ا‬ ‫ال‬ ‫ا‬ --- ‫ط-ب‬ (‫ط‬ Students (plural, masculine)
‫ت‬ + ‫ا‬ ‫ال‬ ‫ا‬ ‫ات‬ ‫ط‬ (‫ط‬ Students (plural, feminine)
+ ‫ا‬ ‫ال‬ ‫ا‬ --- ‫ط‬ (‫ط‬ Student (Singular, masculine)
/ + ‫ا‬ ‫ال‬ ‫ا‬ ‫ه‬ ‫ط‬ (‫ط‬ Student (Singular, feminine)
(+ ‫ي‬ --- --- (‫ط‬ (‫ط‬ He requests (present tense,
singular, masculine)
(+‫ﺗ‬ ‫ت‬ --- --- (‫ط‬ (‫ط‬ She requests (present tense,
singular, feminine)

5
Vowels
English, like the Arabic language, uses vowels and consonants. However, there are two types of
vowels in Arabic, short vowels (symbols) and long vowels (letters). Short vowels are symbols
added above or below the letters (also known as diacritics). These symbols have grammatical
functions to express a word’s grammatical case [13]. Therefore, these symbols can change the
whole meaning of the word. For instance, the word “ ُ2” means “love”, but changing its diacritics
to “ َ2” changes the meaning to “seed”.
Arab people use these diacritics when they speak but to increase the speed of reading and writing,
they often omit them from Modern Arabic written forms. However, omitting diacritics from
words might confuse the reader when these words are read in isolation or out of context. For
instance, the word “, "” alone and without diacritics might have many meanings, such as “eye”,
“spy”, “assistance”, “a spring of water” or “‫”ع‬ (one of the Arabic letters). When the word “, "”
has its diacritics, or is in context without diacritics, this confusion will be resolved. In rare cases,
the reader will not be able to differentiate between un-vocalised words even if they are put in their
context. For instance, in the sentence “ ( 4 ‫ن‬ "” the reader will not be able to know if the writer
intended “Amman is nice” or “Oman is nice”, because the word “‫ن‬ "” without diacritics might
mean “Amman” or “Oman”. To resolve this confusion, the writer has to put the suitable diacritics
on the word (to be “‫ن‬ ّ َ"”, which means “Amman”, or “‫ن‬ َ ُ"”, which means “Oman”).
Keeping the diacritics might be useful in applications such as translation and transliteration in
order to overcome the aforementioned confusion problems and to detect words’ grammatical
cases. In contrast, keeping the diacritics in other applications might not be necessary, such as in
stemming applications that return a word to its original root and in CAs that use Modern Arabic.
Existing the diacritics will increase the processing/analysing complexity to the maximum in
Arabic. In contrast, there is no diacritics in English which leads to reduce the analysing
complexity and save the computational processing time and effort.
Long vowels in Arabic are comprised of three different letters: ‫,ا‬‫و‬ and ‫.ي‬ These letters can either
be originally part of the root or are infixes in a word. Long vowels might change or replace each
other during a word transformation process. For instance, the word “‫م‬ 8” means “he stands” and
the word “‫م‬ 9 ” means “he is standing”, both of which have the same root “‫م‬ 8”. This example
shows that the vowel letter “‫”و‬ in the root “‫م‬ 8” has been replaced with the vowel letter “‫”ا‬ during
the transformation process of word “‫م‬ 8”. No fixed rules control these changes during a word
transformation [21], which increases the difficulty of analysing such words to understand the
sentence by an NLP based CA.
Nouns and verbs
A noun is “a word that indicates a meaning by itself without being connected with the notion of
time” [22], whilst a verb is a word that has meaning with being connected with notion of time
[16].
Unlike English, nouns in Arabic express three number-based cases: singular, dual and plural. For
instance, the word “teacher” in English expresses the singular form of the noun and “teachers”
expresses dual or plural cases. In contrast, the Arabic word “‫رس‬ “ meaning “a teacher” express
the singular status of the noun, and the words ‫ن‬ ‫ر‬ and , ‫ر‬ meaning “two teachers” express
the dual case; ‫ن‬ ‫ر‬ and , ‫ر‬ meaning “three teachers or more” expresses plural status. This
happened by adding some affixes to the mentioned word. In dual and plural nouns, as shown in
the previous example, two different words are used to express their cases. The selection process
between the two words for each pair is based on complicated grammatical rules or using

6
diacritics. Therefore, Arab people habitually make mistakes when choosing between them. This
will make detect and analyse these words variations for a CA understanding is much difficult.
Most natural languages differentiate between male (masculine) and female (feminine) words. In
Arabic, this differentiation is expressed simply by pronouns such as “ ‫”ھ‬ “he” and “;‫”ھ‬ “she”. In
addition, this differentiation might be achieved by adding some affixes to words to express their
gender. For instance, in English, the word “player” expresses both male and female, but in Arabic
the word “ " “ is used to express the male player, and the suffix letter “‫“ه‬ should be added at the
end of the word “ " “ to be “/ " ” for a female player.
In Arabic, masculine and feminine nouns might be related to humans, animals, or even inanimate
objects. In English, all inanimate objects use the word “it” to express them, but in Arabic the
inanimate object can be verbally masculine or feminine. For instance, the word “table” is
considered as feminine “ ‫و‬ ‫ط‬ ;‫“ھ‬ and the word “chair” is considered masculine “; $ ‫.“ھ‬
Therefore, detecting such pronouns (“He” “ ‫,”ھ‬ “She” “;‫)”ھ‬ in Arabic does not reflect if the
intended something is a human or an inanimate. This will increase the challenge to detect and
analyse a sentence in Arabic.
Verbs in Arabic can be generated to express masculine and feminine. For instance, the verb
“‫رس‬ ” “he is studying” is assigned for a male student, while “‫رس‬ ‫”ﺗ‬ “she is studying” assigned
for a female student. Moreover, unlike the English, verbs can express the number of people who
are doing the action. For instance, the verb “‫رس‬ ” “he is teaching” is used to express that there is
one male person doing the action of teaching, while the verb “‫ن‬ ‫ر‬ ” “they are teaching” is used
to express that there are two male people doing this action. However, the verb “‫ن‬ ‫ر‬ ” “they are
teaching” is used to express that there are more than two male teachers doing the action. This
richness of Arabic morphology makes it difficult to an NLP based CA to analyse such verbs for
the large number of affixes and then inability of determining if the intended person is single, dual
or plural.
Proper nouns and foreign Words
Proper nouns in Arabic include the “name of a specific person, place, organization, thing, idea,
event, date, time, or other entity” [23]. Unlike English, Arabic does not support capitalisation of
proper nouns such as “ < ” “Mohammad” [24]. Differentiating the proper nouns in English is
easy as the word has been written capitalised. In contrast, in Arabic there is no way to
differentiate them. In addition, foreign arabised words are words that are borrowed from other
languages, mainly from English, such as microwave”= ‫$و‬ ”. These words cannot be broken
down into roots because originally they are not Arabic words and thus do not have roots. As a
result, a Conversational Agent must deal with such a word as it is. Arabised words have no
standardised writing form. For instance, the word “microwave” has many written forms, such as
“= ‫$و‬ ” and “= ‫$و‬ ”. Thus the challenge of dealing with these words is recognising them
within the conversation. This increases the difficulty of detecting such nouns in textual Arabic
conversations.
Unfortunately, NLP suffers from many limitations and it still faces many challenges. For
instance, a word might have more than one meaning (Lexical ambiguity) or a sentence might
have more than one structure (structure ambiguity). In addition, these ambiguities are different
between languages. A language has one phrase to express something while another language
might have a single word to express the same thing. Moreover, the real life conversations are
often grammatically incorrect and the conversation processing in the machine takes a long time
[25][26]. Furthermore, the natural language used by humans is extremely rich in form and
structure and very ambiguous.

7
These challenges increase when developing a CA for rich inflectional and derivational language
such as Arabic. An efficient NLP based CA should interpret the utterance, determine and perform
the actions that must be taken as response to the utterance [27]. For instance, if a user entered “I
would like to buy it now”. The CA should determine the meaning of the utterance that he/she
wants to buy something maybe it mentioned in earlier conversation or it is out of the conversation
context. Once the CA interpreted the utterance, it should know how to acts on this utterance such
as continuing the purchase procedures. Finally, the CA should respond with appropriate actions
that might include presenting information or accessing/writing to a database [27].
In the last few years, the Arabic based computational processing tools has gained a more
concentration due to the large amount of the online Arabic available data [28]. These tools can be
applied in different areas such as information retrieval, speech recognition, machine translation,
localization, sentiment analysis, CAs and tutoring systems. The need of providing a user with
high quality tools is important to keep up with the huge amount of Arabic data growth [29].
However, the Arabic language is still young in the NLP area in general [30].
An Arabic CA based upon NLP will face all of the above discussed challenges.The Arabic
language has the flexibility of sentence structuring rather than the English. The sentence structure
in Arabic has three forms [13], which are: (from right to left): [object][subject][verb] ( ‫أ‬<
2 # ‫”)ا‬Mohammad eats an apple”, [object][ verb][ subject] ( 2 # ‫ا‬ ‫أ‬ < ) ”Mohammad eats an
apple”, and [verb][subject][ object] ( < ( ‫أ‬ 2 # ‫)ا‬ ”the apple has been eaten by Mohammad”.In
contrast, the sentence structure in English might be [subject][verb][object] (‘Mohammad eats an
apple’). Consequently, this flexibility of sentence structuring in Arabic will increase the
complexity of building a CA based on NLP in terms of actual sentence understanding. In Arabic
language, the research into computational semantics is much smaller than other areas in NLP,
because of its higher complexity [13]. According to [31], “There seems to be no agreement on an
efficient way for having an adequate morphological analysis/generation”. Moreover, the NLP
approach is expensive computationally. In contrast, in the English language, these challenges is
much less in its complexity as the sentence’s structure in English is simple as illustrated
above.Another limitation that Arabic might faces is that NLP tools has been developed for
western languages so it is not easy to adapt with Arabic language as it has different features [12].
In addition, the transfer of knowledge and technology as most recent publications in science and
technology is from the English language.
Understanding the sentence is the most efficient way to handle conversations either in Arabic or
in English. However, this computing based understanding of the sentence is still the dream for the
researchers as it needs further research. Apparently, this dream is much longer for the Arabic
language NLP researchers [31]. Even the Machine Translation (MT) from Arabic to English is a
not easy task as it mainly based on understanding the Arabic sentence [32]. In Arabic, the
sentence is long; its length is 20 to 30 words and sometimes exceeds the 100 words [32].
According to [32], “The Arabic sentence is complex and syntactically ambiguous due to the
frequent usage of grammatical relations, order of words and phrases, conjunctions, and other
constructions. Consequently, most of the researches in Arabic MT mainly concentrated on the
translation from English to Arabic.”. A few number of CAs based on NLP has been developed
such as the GALAXY [33] which based on the English language and no similar related work has
been reported for the Arabic.
Finally, when Arabic language computing based research progress to reach understanding the
sentence, the NLP approach might be best way to build an Arabic CA.

8
2.2 Arabic Pattern Matching based Conversational Agent
Text-based Pattern Matching (PM) in computer science is the process of searching for a string or
sequence of strings in a piece of text to find all occurrences of these strings inside that text. PM
has been used in a large number of applications [34], such as computer parsers [35], spam filters
[36], natural language processing [37], computational molecular biology [38], question answering
systems [39] and CAs [3]. From a CA perspective, PM is a mechanism that uses an algorithm to
handle user conversations by matching CA’s patterns against a user’s utterance [40]. A typical
pattern consists of a collection of words, spaces, and wildcards. A wildcard is a symbol used to
match a portion of the user’s utterance.
PM algorithms have been successfully used in the development of a number of CAs such as
InfoChat[6, 41] , ArabChat[9] and ALICE [3]. These CAs organise their scripted domain into
contexts, each of which has a number of related rules. Each rule has structured patterns that
represent a user’s utterance. In addition, a rule has associated response to reply to the utterance. In
the CA’s engine that rely on the PM approach, it match the user’s utterance with the scripted
patterns. These patterns distributed in rules which themselves represents different topics. The
matched rule will be fired and its response will send back to the user.
PM is considered one of the most successful methods for developing CAs that show or at least
give the impression of some kind of intelligence [42]. The PM technique has many advantages
compared with other techniques, including that it is easy to understand and it is language
independent which it might be suitable for English and Arabic. In addition, PM based CAs do not
require complex pre-processing stages such as “part of speech tagging” that require extra time to
process and thus the PM is not expensive computationally. Consequently, CAs developed using
PM can support conversations effectively for large numbers of users in a real-time environment
like the Internet [42]. Moreover, it resolves many linguistic challenges that NLP technique faces,
such as morphological changes occurring on a word through adding affixes to it. All of these
morphological changes can be covered using pattern’s wildcards. It is easy to resolve
grammatical and spelling errors in a user’s utterance using the pattern wildcards. Resolving
spelling and grammatical errors is an important function that a CA needs to keep the conversation
going, especially as research studies show that the amount of grammatical and spelling errors
increases as conversations progresses [43].
The main disadvantage of a PM based CAs is that it requires a large number of patterns to
implement a coherent domain. This large number of patterns comes from many issues. Firstly,
there are many diverse ways that a user can construct his/her utterance (using different sentence
forms or using different synonyms). Secondly, the natural language morphological richness forces
a scripter to cover all expected keyword’s changes to meet different cases of the scripted keyword
such as singular/plural.
This disadvantage will be faced by both languages (English and Arabic). In Arabic, this challenge
is bigger as it rich in flectional and derivational features as described in the previous section. The
word transformation in Arabic is much large and complex than the English specially when adding
infixes. Although, using the PM technique for Arabic language is extremely affected by the
affixation features (which increase the pattern scripting effort), it is possible to alleviate the
patterns scripting difficulties by using different techniques such as stemming [44]. Using the
stemming technique as a pre-processing stage, a CA can eliminate a word from its affixes to
result its root to be used in patterns scripting instead of covering many variations o f a word.
However, this stemming technique might increase the computational cost but it can decrease the
number of scripted patterns and thus decreasing the computational cost as well.

9
2.3 Sentence Similarity Measures based CA
Generally, similarity as a term is used to describe the similarity level among two objects.
Similarity has been debated in number of fields such as philosophical, linguistic, and information
theory [45]. Semantics is “the study of the meaning of linguistic expressions” [13]. In computing,
generally, and in artificial intelligence, especially, similarity-based research has been conducted
in many applications such as information retrieval, information extraction, machine translation,
question answering, and conversational agents [46].In this paper, it will be focused on sentence
similarity measures as an approach to build the CA.
Sentence similarity techniques try to measure the semantic similarity level between sentences. A
sentence is a collection of words. Therefore, a semantic sentence similarity is related to those
words semantic similarity [46].
Two main approaches have been used to measure the similarity between sentences are “Latent
Semantic Analysis” (LSA) [47] and”sentence similarity based on semantic networks and corpus
statistics” (SSSN) [48].
LSA analyses a large corpus of a text’s words using statistical computations and then generates a
representation that captures the similarity of words and text passages [47]. A matrix of words is
generated based on the number of times that a word appeared in a specific context without
consideration for the order of words in a sentence [47]. Then, LSA uses the Singular Value
Decomposition (SVD) technique to decompose the words matrix to reduce its size. SVD is an
analysing technique that tries to reduce the dimensional representation of the matrix by keeping
the entries that have the strongest relationship between words and their occurrences in sentences.
However, LSA does not deal with syntactic relations (words order is not important) and with
polysemy (words with multiple meanings). Consequently, this causes problems that might lead to
the inability to correctly analyse the sentences. In addition, the SVD technique is expensive
computationally.
The SSSN approach can overcome the LSA’s limitations [49]. SSSN is a technique based on
computing the similarity through the use of lexical/semantic resources such as a WordNet [50].
WordNet is a machine-readable lexical database developed at Princeton University. WordNet has
four word categories (nouns, verbs, adjectives, and adverbs). WordNet’s words are grouped into
synonyms called synsets. Synsets are connected by means of semantic and lexical relations.
However, Most of the published work in sentence similarity measures has been done for the
English language [51]. The WordNet research work for Arabic is a young area as it has been
started in 2006 with the collaboration of several companies and universities [52][31]. Where, the
SSM research work for Arabic language has been started in 2012 according to [51] with creating
of an Arabic benchmark dataset of only 70 word pairs. As a result, the sentence similarity area in
Arabic is still young and needs further research to rely on it for CA building.
As discussed in the previous section that pattern matching suffering from the big number of
needed patterns in order to build an robust CA. Sentence similarity approach came as an
alternative way of building a CA to overcome PM technique challenge.
According to [53-56], applying the sentence similarity technique in CA building is more effective
in terms of reducing the scripting effort to its minimum. The Sentence similarity technique
replaces groups of scripted patterns into a few natural language sentences, which leads to the
reduction of the scripting time and the effort of pattern maintenance [55]. For instance, table 2
[49] shows that the scripting differences in a rule handle the same topic between the PM and
sentence similarity techniques. Where ‘P’ is a pattern expression, ‘S’ represents a sentence, and
‘R’ represents the rule’s response.

10
Table 2: shows an example of scripting differences in a rule handle the same topic between the PM and
sentence similarity techniques (taken from [49]).
A Rule scripts using Pattern Matching A rule scripts using sentence similarity
<Rule 1>
P: * money
P: * cash
P: * funding
.
.
R: I’m sorry to hear that
<Rule 1>
S: I have no money
R: I’m sorry to hear that
However, the measurement of similarity among sentences is an “uncertainty problem” because
the real similarity process from the human judge might depend on information factors which are
usually related to time and a situation.[46]. Although, capturing these hidden information factors,
they are a big challenge as the current research cannot totally determine how the human brain
works and how he/she capture such these information factors [46]. Moreover, no research in [49,
55, 56] have mentioned details as to if they tried to evaluate their CAs in an online environment
like in terms of the elapsed time needed to handle an user’s utterance.
In Arabic language, the research into computational semantics much smaller than other areas in
NLP perhaps due to its higher complexity [13]. In addition, a useful database such as “WordNet”
which might help the CA researcher to use it in order to build their CAs based on sentence
similarity, is still a young area. Arabic WordNet (AWN) have been started in 2006 as mentioned
above.
The Arabic CA based on a sentence similarity approach will face many challenges such as Arab
people use three types of Arabic language (Classical, Modern and Colloquial) which might make
their utterances mixed between them. Thus, increasing the need for a CA to deal with colloquial
words as an example. Moreover, some Arabic words have different semantic meanings for
instance, the word (, ") alone and without diacritics might have many meanings such as “eye”,
“spy”, “assistance”, “a spring of water” or “‫”ع‬ which is one of the Arabic letters. The word “?@ ”
could means “river side” or it means a “financial institution”. This variety of meaning for the
same word could cause “WordNet” to misclassify such a word [31]. In addition, Arabic has
neither capital letters nor acronyms in order to detect proper names such as names of people,
names of months, and names of cities [31]. Moreover, some Arabic names increase the ambiguity
of classifying them because they are sharing some animal, months and days names. For instance,
the word “/ 4” might mean day of the week which is “Friday” and might be a male person name
and the word “A ” might means an animal name which is “Lion” or might means a male person
name. The word “‫ن‬ B ‫”ر‬ might mean an Arabic month’s name which is “Ramadan” or might
means male person name.
3. CONCLUSION
In this paper, a brief description for Arabic language has been introduced with the comparing to
the English language. This paper focused on the Arabic language complexities and its challenges
in computing based work in general. In addition, this research work has determined its scope by
discussing the Conversational Agent (linguistic text-based) and its building approaches. Three

11
main approaches for CA building which are; NLP, SSM and PM. Each approach has advantages
and disadvantages in building the CAs. This paper debated briefly the three approaches
challenges and limitations in building the CA for the Arabic language comparing to the English
language. It has been released that the Arabic language is more challengeable than the English in
terms of a CA building for many general reasons. Firstly, the complexity of the Arabic language
itself as it is a Semitic language with a high rich derivational and inflectional features.This
complexity derived from different aspects as this paper described. Secondly, most of the research
work in the NLP and SSM areas were for English language. This increased the challenge as it
needs a further research for those areas before start building a CA.
ACKNOWLEDGMENT
The authors acknowledge the financial support that received from the Applied Science University
in Jordan to present this work.
REFERENCES
[1] Turing, A., Computing machinery and intelligence. Mind, 1950: p. 433-460.
[2] Weizenbaum, J., ELIZA: A computer program for the study of natural language communication
between man and machine. Communications of the ACM., 1966. Vol 10.: p. PP 36-45.
[3] Wallace, R.S. ALICE: Artificial Intelligence Foundation Inc. 2008 [cited; Available from:
http://www.alicebot.org.
[4] ConvAgent. ADAM Conversational Agent Demo. 2001 [cited 23-02-2011]; Available from:
http://www.convagent.com/convagent/adam3.aspx.
[5] Kopp, S., et al., A conversational agent as museum guide: design and evaluation of a real-world
application, in Lecture Notes in Computer Science. 2005, Springer-Verlag. p. 329-343.
[6] Sammut, C. and D. Michie, InfochatTM Scripter’s Manual, Convagent Ltd. 2001: Manchester.
[7] Kadous, M. and C. Sammut, InCA: A Mobile Conversational Agent. Trends in Artificial Intelligence,
2004. Volume 3157: p. pp 644-653.
[8] Alobaidi, O., et al. Abdullah: An Intelligent Arabic Conversational Tutoring System for Modern
Islamic Education. in WCE 2013. 2013. London, U.K.
[9] Hijjawi, M., et al. ArabChat: An Arabic Conversational Agent. in Computer Science and Information
Technology (CSIT), 2014 6th International Conference on. 2014. Amman, Jordan.
[10] UN. United Nations. 2011 [cited; Available from: http://www.un.org.
[11] Ryding, K., A Reference Grammar of Modern Standard Arabic. 2005: Cambridge university press.
[12] Ali, F. and S. Khaled, Arabic Natural Language Processing: Challenges and Solutions. 2009. 8(4): p.
1-22.
[13] Habash, N., Introduction to Arabic Natural Language Processing, ed. U.o.T. Graeme Hirst. 2010:
Morgan & Claypool.
[14] Al-Saidat, E.and I. Al-Momani, Future markers in modern standard arabic and jordanian arabic: A
contrastive study. European Journal of Social Sciences, 2010. 12.
[15] Al Ameed, H., et al., Arabic light stemmer: A new enhanced approach, in The Second International
Conference on Innovations in Information Technology (IIT’05). 2005.
[16] Abu-Chacra, F., Arabic: An Essential Grammar. 2007.
[17] Crystal, D., Dictionary of linguistics and phonetics. Sixth Edition ed. 2008: Blackwell.
[18] Khoury, R., F.Karray, and M. Kamel, Keyword extraction rules based on a part-of-speech hierarchy.
International Journal of Advanced Media and Communications, 2008. V2: p. pp 138-153.
[19] Al-Fedaghi, S. and F. Al-Anzi. A New Algorithm to Extract Arabic Root-Pattern Forms. in
Proceedings of the 11th National Computer Conference. 1989. Saudi Arabia.
[20] Ahmed, A.R. and F.S. Khaled, Lexical analysis of inflected Arabic words using exhaustive search of
an augmented transition network. Softw. Pract. Exper., 1993. 23(6): p. 567-588.
[21] Al-Shammari, E., Improving Arabic text processing via stemming with application to text mining and
Web retrieval, in Graduate Faculty. 2010, George Mason University. p. 201.
[22] Abuleil, S. and M. Evens, Extracting an Arabic Lexicon from Arabic Newspaper Text. Computers
and the Humanities, 2002. 36(2): p. 191-221.

12
[23] Abuleil, S., K. Alsamara, and M. Evens, Acquisition system for Arabic noun morphology, in
Proceedings of the ACL-02 workshop on Computational approaches to semitic languages. 2002,
Association for Computational Linguistics: Philadelphia, Pennsylvania.
[24] Diab, M., K. Hacioglu, and D. Jurafsky, Automatic tagging of Arabic text: from raw text to base
phrase chunks, in Proceedings of HLT-NAACL 2004: Short Papers. 2004, Association for
Computational Linguistics: Boston, Massachusetts.
[25] Fernezelyi, M., Z. Szegedy, and L. Robert, Smalltalk: interactive installation, in Proceedings of the
14th annual ACM international conference on Multimedia. 2006, ACM: Santa Barbara, CA, USA.
[26] Michie, D., Return of the Imitation Game. Linköping Electronic Articles in Computer and
Information Science, 2001(Vol. 6).
[27] Lester, J., K. Branting, and B. Mott, Conversational Agents, in The Practical Handbook of Internet
Computing M. Singh, Editor. 2004, Chapman and Hall/CRC. p. 1144.
[28] Abdelali, A., J. Cowie, and S. Soliman. Building A Modern Standard Arabic Corpus. in Workshop on
Computational Modeling of Lexical Acquisation. 2005. Croatia.
[29] ElHadj, Y., I. Al-Sughayeir, and A. Al-Ansari. Arabic part-of-speech tagging using the sentence
structure. in Proc. Second International Conference on Arabic Language Resources and Tools 2009.
[30] Shaalan, K., Rule-based Approach in Arabic Natural Language Processing. 2010.
[31] Black, W., et al., Building aWordNet for Arabic, in In Proceedings of The fifth international
conference on Language Resources and Evaluation (LREC 2006). 2006: Italy.
[32] Shaalan, K., Machine Translation of Arabic Interrogative Sentence into English. National Agricultural
Research Information Managment System, 2000.
[33] Seneff, S., et al. (1998) GALAXY: A reference architecture for conversational system development.
Volume,
[34] Sedgewick, R. and K. Wayne, Algorithms. 4th edition ed. 2011.
[35] Alessandro, W. and P. Ian, OMeta: an object-oriented language for pattern matching, in Proceedings
of the 2007 symposium on Dynamic languages. 2007, ACM: Montreal, Quebec, Canada.
[36] Freschi, V., A. Seraghiti, and A. Bogliolo. Filtering Obfuscated Email Spam by means of Phonetic
String Matching. in In Proceedings of ECIR. 2006.
[37] David, D.L., S. Karen, and J. rck, Natural language processing for information retrieval. Commun.
ACM, 1996. 39(1): p. 92-101.
[38] Rouchka, E., Pattern Matching Techniques and Their Applications to Computational Molecular
Biology - A Review, in WUCS. 1999.
[39] Hang, C., K. Min-Yen, and C. Tat-Seng, Soft pattern matching models for definitional question
answering. ACM Trans. Inf. Syst., 2007. 25(2): p. 8.
[40] Pudner, K., K. Crockett, and Z. Bandar, An Intelligent Conversational Agent Approach to Extracting
Queries from Natural Language, in World Congress on Engineering, International Conference of Data
Mining and Knowledge Engineering. 2007. p. pp 305-310.
[41] ConvAgent. ConvAgent Foundation- ADAM Conversational Agent. 2011 [cited 20-02-2011];
Available from: www.ConvAgent.com.
[42] Timothy, B. and G. Toni, Health dialog systems for patients and consumers. J. of Biomedical
Informatics, 2006. 39(5): p. 556-571.
[43] Maragoudakisa, M., et al. Natural language in dialogue systems. A case study on a medical
application. in Proceedings of Panhellenic Conference with International Participation in Human–
Computer Interaction. 2001. Patras, Greece.
[44] Hijjawi, M., et al., An Application of Pattern Matching Stemmer in Arabic Dialogue System, in 5th
International KES Conference on Agents and Multi-agent Systems – Technologies and Applications.
2011: Manchester.
[45] Hatzivassiloglou, V., J. Klavans, and E. Eskin. Detecting text similarity over short passages:
Exploring linguistic feature combinations via machine learning. in In Proceedings of the Conference
on Empirical Methods in Natural Language Processing and Very Large Corpora. 1999.
[46] Feng, J., Y. Zhou, and T. Martin, Sentence Similarity based on Relevance, in Information Processing
and Managment of Uncertainty in Knowledge-Based Systems. 2008: Spain.
[47] Landauer, T., P. Foltz, and D. Laham, Introduction to Latent Semantic Analysis. Discourrse
Processes, 1998. V 25: p. pp 259-284.
[48] Li, Y., et al., Sentence similarity Based on Semantic Nets and Corpus Statistics. IEEE Transactions on
Knowledge and Data Engineering, 2006. V 18(8): p. 1138-1149.
[49] O’Shea, K., Z. Bandar, and K. Crockett, A Conversational Agent Framework using Semantic
Analysis. International Journal of Intelligent Computing Research, 2010. V 1(1).

13
[50] George, A.M., WordNet: a lexical database for English. Commun. ACM, 1995. 38(11): p. 39-41.
[51] Almarsoomi, F., et al., ArabicWord Semantic Similarity. World Academy of Science, Engineering
and Technology2012. Vol:6.
[52] Black, W., et al., Introducing the Arabic WordNet Project, in In Proceedings of the Third
International WordNet Conference. 2006: Korea.
[53] O'Shea, J., et al., A comparative study of two short text semantic similarity measures, in Proceedings
of the 2nd KES International conference on Agent and multi-agent systems: technologies and
applications. 2008, Springer-Verlag: Incheon, Korea.
[54] Li, Y., et al., A Method for Measuring Sentence Similarity and its Application to Conversational
Agents, in FLAIRS. 2004.
[55] Karen O’Shea, Z.B., and Keeley Crockett, A Novel Approach for Constructing Conversational
Agents using Sentence Similarity Measures. 2008.
[56] O'Shea, K., Z. Bandar, and K. Crockett, Towards a New Generation of Conversational Agents Based
on Sentence Similarity, in Advances in Electrical Engineering and Computational Science. 2009,
Springer Netherlands. p. 505-514.
Authors
Mohammad Hijjawi is currently an assistant professor at the Applied Science
University in Amman, Jordan. Dr.Hijjawi holds a Ph.D. degree in Computer Science
from Manchester Metropolitan University, UK. His research interests include:
Conversational Agents, machine learning, Data Mining and Arabic language
processing.
Yousef Elsheikh is an assistant professor of Information Technology at the Applied
Science University. He holds PhD in Informatics from University of Bradford, UK and
MSc in Information Technology from University of the West of England, UK. He is
currently working as a head of the Computer Information Systems department at the
Applied Science University. His research interests includes conceptual modeling,
ebusiness applications, information systems engineering, knowledge based
representations, Ontologies and issues in software engineering.

ARABIC LANGUAGE CHALLENGES IN TEXT BASED CONVERSATIONAL AGENTS COMPARED TO THE ENGLISH LANGUAGE

More Related Content

What's hot (18)

Viewers also liked (20)

Similar to ARABIC LANGUAGE CHALLENGES IN TEXT BASED CONVERSATIONAL AGENTS COMPARED TO THE ENGLISH LANGUAGE (20)

ARABIC LANGUAGE CHALLENGES IN TEXT BASED CONVERSATIONAL AGENTS COMPARED TO THE ENGLISH LANGUAGE