Experimenting Machine-Learning Algorithms for Morphological

Disambiguation of Arabic Texts

Bilel Elayeb

1,2 a

, Mohamed Firas Ettih

and Raja Ayed

2,4 c

Liwa College of Technology, P.O. Box 41009, Abu Dhabi, U.A.E.

RIADI Research Laboratory, ENSI, Manouba University, Tunisia

Université Paris-Est Créteil, Paris 12 Val de Marne, France

Faculty of Economics and Management of Nabeul, Carthage University, Tunisia

Keywords: Morphological Disambiguation, Arabic Text, Machine-Learning Algorithms, Data Transformation,

Morphological Feature, Classification.

Abstract: Arabic language is characterized by its complexity and its morphological and orthographic variations

including syntactic and semantic diversity of a word. This specificity may cause Arabic morphological

ambiguity. We present in this paper a new architecture for morphological disambiguation of Arabic texts. The

latter can be treated as a classification problem where the set of morphological features’ values represent

classes, and a classification algorithm is used to assign a class to each word’s occurrence based on the context.

The first step consists of identifying the correct morphological analysis of a non-vocalized Arabic word using

the morphological dependencies extracted from the corpus of vocalized texts. Then, we propose a method of

transforming imperfect training datasets into perfect data having precise attributes and certain classes. We

experiment this architecture on a set of machine-learning classifiers using a corpus of classic Arabic texts.

Results highlight some statistically significant improvement of SVM and Naïve Bayes classifiers in terms of

disambiguation rate.

1 INTRODUCTION

Morphology is the field that studies how the smallest

meaningful units, called morphemes, combine to

form lemmas which are the autonomous units

forming the language lexicon. The morphology

systems are useful to support NLP tools such as

analyzers and information retrieval systems (IRS).

Morphological analysis only reveals the various

potential of words’ vowels in a text and interprets

their structures. However, a morphological analyzer

can be used to display all forms of verbs in Arabic. It

can also display multiple forms if the user chooses to

specify not only the root but other morphological

attributes like gender, number, and mode.

Morphological analyzer displays all possible values

of these attributes. It can analyze any form of a given

word with a certain coverage rate.

https://orcid.org/0000-0002-5050-2522

https://orcid.org/0000-0002-2237-7146

https://orcid.org/0000-0003-3339-8388

Numerous applications in the field of Arabic NLP

must deal with the complex morphology of this

language (Elayeb and Bounhas, 2016; Elayeb, 2019).

Morphological analysis is an important step in

automatic speech recognition, Arabic texts

phonetization and automatic Arabic text

summarization (Elayeb et al., 2020). Besides,

information retrieval applications should index

documents and extract relevant characteristics of their

significant entities (Elayeb, 2009; Elayeb et al., 2009;

Bounhas et al., 2011b). Indeed, Information Retrieval

and Knowledge Extraction Systems (IRKES) require

the recognition of useful entities in texts such as

words, phrases, and concepts. The basic level

concerns the word’s structure, i.e. the morphological

level. Indeed, a given word can have several

morphological interpretations, which makes it

ambiguous. For example, the word "

ﻞﻛﺃ

" can be

interpreted as a noun meaning "food" (

ﻞ

ﻛ

ﺃ

; >ak°luN )

Elayeb, B., Ettih, M. and Ayed, R.

Experimenting Machine-Learning Algorithms for Morphological Disambiguation of Arabic Texts.

DOI: 10.5220/0010917300003116

In Proceedings of the 14th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2022) - Volume 3, pages 851-862

ISBN: 978-989-758-547-0; ISSN: 2184-433X

851

and as a verb meaning "to eat" (

ﻞ

ﻛ

ﺃ; >akala). This

phenomenon represents a challenge for

morphologically rich languages such as Arabic (Diab

et al., 2004). Thus, a non-vocalized Arabic word can

have more than 12 interpretations (Habash and

Rambow, 2007; Habash et al., 2009b).

The information retrieval process begins with an

analysis and pre-processing step that aims to index

documents and extract their knowledge (Elayeb,

2009, 2018). Indexing is an essential phase, aiming at

ensuring that documents and queries are represented

by keywords. These terms are standardized forms

obtained by a morphological analysis of words

representative of documents and queries. Arabic

language is characterized by its morphological and

orthographic variations including syntactic and

semantic diversity of a word. This morphological

richness causes ambiguity and difficulty in

identifying the appropriate analysis, which can affect

the definition of indexing terms and change the

queries’ meanings. Adding short vowels to a word

can decrease its ambiguity but does not eliminate it.

Besides, morphological disambiguation is essential to

facilitate and improve the IR indexing task.

Existing Arabic morphological disambiguation

approaches mainly disambiguate the Part-Of-Speech

(POS) feature of words and only cover a particular

type of texts, namely modern Arabic texts. POS

disambiguation consists of determining the

grammatical category of a word in a particular

context. It can also be considered as a classification

problem where the set of POS values represent

classes, and a classification method is used to assign

a class to each occurrence of a word based on the

context. One of the important steps in the

disambiguation task is the suitable selection of the

classification method. Supervised automatic

classification methods were applied. They have used

training techniques to learn a classifier from training

sets annotated with the values of the POS class.

We experiment, analyze and compare in this paper

a series of machine-learning algorithms in order to

solve the problem of morphological disambiguation

of Arabic texts using several morphological

attributes. We test these algorithms on the Kunuz

test

collection (Ben Khiroun et al., 2014; Ayed, 2017;

Ayed et al., 2018ab) containing classical vocalized

Arabic texts. This corpus contains Hadith texts

assigned to the Prophet of Islam Mohammad

(PBUH). It has been the subject of several research

works due to its linguistic, semantic, social richness

and its well-organized structure. However, the TREC

http://www.jarir.tn/kunuzcorpus

collections (2001 and 2002) of Arabic newspapers

articles suffer from the absence of vowels in their

texts, which could generate a certain word sense

ambiguity and a difficulty in identifying its POS

feature as well as its function in the sentence.

The remaining of this paper is organized as

follows: existing works on morphological

disambiguation of Arabic texts are summarized and

discussed in Section 2. The main sources of

morphological ambiguity are presented in Section 3.

Section 4 presents the architecture of the proposed

approach as well as the training and testing data

preparation phases. Section 5 introduces the

experimental results as well as a comparative study

between the ML classifiers used in the morphological

disambiguation of Arabic texts. We conclude our

work in Section 6, and we suggest some perspectives

for future research.

2 RELATED WORK

In the literature, disambiguation approaches can be

classified into three categories, namely rule-based

approaches, statistical approaches, and their

combination known as hybrid approaches. We briefly

detail and discuss these works in this section.

Rule-based or linguistic approaches have used a

knowledge base of rules proposed by linguists to

assign labels to different morphological attributes.

Systems dealing with linguistic disambiguation

approaches are described by (Othman et al., 2004;

Daoud, 2009). These techniques are based on

heuristics, contextual and non-contextual rules. These

rules are often classified into grammatical, structural,

and logical categories (Al-Ansary, 2005). Daoud and

Daoud (2009) have proposed a specific type of

analyzer called En-Converters written in UNL

(Universal Networking Language) and EnCo

which

is a rule-based programming language. The authors

have defined several types of disambiguation rules

combining morphological and syntactic contextual

dependencies. However, the authors did not perform

any experimental evaluation, making it difficult to

assess their approach in terms of coverage,

reusability, and precision. Some researchers, such as

(Othman et al., 2004), have exploited full syntactic

parsing for morphological disambiguation.

Statistical approaches are based on learning

models from annotated corpora. They incorporate (i)

statistical models such as the HMM (Hidden Markov

Model) in which the modelled system is assumed to

http://libraries.unl.edu/

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

852

be a Markovian process of unknown parameters or

(ii) classification methods such as SVM to compute

probabilities of each value resulting from a given

word POS. A model can be used to automatically

classify other texts by referring to the already

calculated probabilities.

For example, Mansour et al. (2007) have

combined probabilities calculated on Arabic and

Hebrew learning sets to classify the words POS in

Arabic texts. ElHadj et al. (2009) have presented a

POS tagger system that combines morphological

analysis and the Markov model. The tagging process

is based on the Arabic sentence structure. First, the

text is fully morphologically analyzed to reduce the

number of possible POS values. Second, the HMM

statistical model, based on the structure of the Arabic

sentence, is used to assign each word the exact POS

value. ElHadj et al. (2009) have used their annotated

corpus which is composed of classical Arabic books.

The total of words in this corpus is approximately

21.000 words. Diab et al. (2004) have developed a

morphological classifier using SVM. They have

trained and tested the classifier using an Arabic

Treebank of 4.000 training sentences and 100 test

sentences. The most widely used tool, denoted

MADA, is developed by (Habash and Rambow,

2005) to solve the morphological ambiguity of Arabic

texts. This tool achieves over 86% accuracy in

anticipation of diacritics. MADA have prioritized

complete analysis in terms of overall accuracy. From

its version 2.1, MADA has used the ARAGEN

version of BAMA (Habash, 2007) which can generate

morphological analysis even for unknown words that

are not covered by the lexicon.

A hybrid approach combines linguistic rules with

statistical information in order to resolve

morphological ambiguity. In (Tlili-Guiassa, 2006),

the author has proposed an approach that analyses

grammatical and inflectional affixes and grammatical

rules based on the MBL (Memory Based Learning)

approach. It is applied to classify a collection of

Quranic and educational texts. Zribi et al. (2006) have

combined the rule-based approach with an HMM

trigram-based tagger. The training of the trigram

classifier has been performed using texts containing

6.000 words. Heuristic rules were applied to select an

analysis among the proposed results.

Khoja (2001) has implemented a hybrid approach

based on the Viterbi algorithm. The proposed

technique computed two probabilities on an

annotated corpus composed of 50.000 words: (i) a

lexical probability, which is the probability that a

word has a certain value of a morphological attribute;

and (ii) a contextual probability, which is the

probability that one tag precedes or succeeds another.

A list of grammatical rules is prepared from these

statistics in order to achieve an accuracy greater than

90%.

Belguith and Chaâben (2006) have proposed a

method for Arabic morphological analysis and

disambiguation. It is classified as a statistical

approach, but it also includes rules. This approach is

based on five steps: (i) segmentation of the text into

words; (ii) morphological pre-processing which

consists in removing the clitics based on a predefined

list; (iii) affixal analysis which recognizes the basic

elements of a word, namely the root and the affixes;

(iv) morphological analysis based on MORPH2

(Belguith and Chaâben, 2006); and (v) post-

processing which identifies word groupings based on

a lexicon and a set of rules. This approach has

computed the morphological attributes of each word

using a determined vocabulary.

Bousmaha et al. (2016) have suggested a hybrid

disambiguation approach based on diacritics

selection at different levels of analysis. This hybrid

approach has combined a linguistic approach with a

multicriteria decision and could be considered as an

alternative to solve the problem of morpho-lexical

ambiguity. During the assessment process, the

authors have relied on the online morphological

analyzer and obtained encouraging results with an F-

measure of over 80%.

Bounhas et al. (2015a) have proposed three

possibilistic classifiers for Arabic morphological

disambiguation: (i) the first classifier is based on

possibility measure, (ii) the second one is based on

necessity measure and, (iii) the third one is based on

the combination of these two measures. The authors

have enriched these classifiers with the information

gain scores, useful as weights for the classification

attributes, to reduce the required space for resolving

the contextual ambiguity, which simplify the

disambiguation process.

Later, Bounhas et al. (2015b) have suggested a

hybrid possibilistic approach that combines the

possibilistic classifier with linguistic rules to assign

labels to different morphological attributes. This

approach has improved the disambiguation rate of

Arabic texts. The authors have also presented an

approach dealing with "out-of-vocabulary" words

whose morphological analysis is unknown. These

possibilistic and hybrid classifiers were tested, in

terms of disambiguation rate, on the Arabic "Kunuz"

collection and compared to the three ML classifiers

SVM, Naïve Bayes and Decision Tree.

More recently, Ayed et al. (2018) have

investigated possibilistic morphological

Experimenting Machine-Learning Algorithms for Morphological Disambiguation of Arabic Texts

853

disambiguation of structured Hadiths Arabic texts

using semantic knowledge. Training and testing steps

required morphological attributes and have been

performed using AlKhalil analyzer. The authors have

taken advantage of the XML format of the structured

Hadiths texts of "Kunuz" collection to benefit from

the available semantic information. They have

involved semantic attributes in their possibilistic

classifiers. Experimental results showed an

improvement in the disambiguation rates of

possibilistic classifiers when considering semantic

knowledge.

3 ARABIC MORPHOLOGICAL

AMBIGUITY

Arabic words are often ambiguous in their

morphological analysis because of the richness of the

Arabic affixation and clitic systems (Elayeb and

Bounhas, 2016; Elayeb, 2018; Elayeb, 2019; Elayeb,

2021). Besides, the omission of short vowels in

standard orthography can amplify the ambiguity of

certain words. We determine in the following the

main factors of ambiguity in Arabic texts, namely (i)

agglutination ambiguity; (ii) derivational and

inflectional ambiguity; and (iii) ambiguity of non-

vocalized texts.

3.1 Agglutination Ambiguity

In Arabic, some prepositions and pronouns stick to

the nouns, verbs, adjectives and particles to which

they are linked. Agglutination is one of the problems

found in processing the Arabic language since an

agglutinated word can be translated into a sentence in

English. The Arabic word "

ﺎَﻨَﻧﻮ

ﺴ

ﻨ

ﺘ

ﺳ

ﺃ"

(>asatan°suwnanA) can be translated as "Are you

going to forget us?". Then the Arabic language

analyzer must segment this word to recognize the

root. Moreover, sentences in Arabic do not follow an

exact structure such as in French or English: Subject

+ Verb + Complement, which makes it difficult to

process these texts.

3.2 Derivational and Inflectional

Ambiguity

Some grammatical factors such as verbs’ conjugation

or nouns’ declension generate inflection of the words

forms. The word "

ﻥﻮ

ﻤ

ﱠ

ﻠ

ﻜ

ﺘ

ﻳ" (they speak;

yatakal~amuwna) is the result of the concatenation of

the prefix "

ﻱ" (ya) indicating the present tense and

the suffix "

ﻥﻭ" (wna) indicating the masculine plural

of the verb

ﻢ

ﱠ

ﻠ

ﻜ

ﺗ" (he talk; takal~ama). Many

inflectional operations produce a slight change in

pronunciation without an explicit effect on spelling

due to the lack of short vowels. A recurring example

is the ambiguity of active vs passive vs imperative

forms. As an indication, the form "

ﻞﺳﺭﺃ" (>rsl)

becomes "

ﻞ

ﺳ

ﺭ

ﺃ" (he sent; >ar°sala) in the active

voice, "

ﻞ

ﺳ

ﺭ

ﺃ" (he is sent; >ur°sila) in the passive

voice and "

ﻞ

ﺳ

ﺭ

ﺃ" (sends; >ar°sil) in the imperative.

Besides, some affixes can be homographic. For

example, the prefix "ﺕ" (t) can indicate both male or

female person. For example, the word

ﻞ

ﻛ

ﺄ

ﺗ" (ta

>°kulu) can be translated into "she eats" and "you

eat".

The ambiguity is also caused by the derivation.

The Arabic word is the result of a combination of

root, vowels, prefixes, infixes, suffixes and a

morphological scheme. Prefixes and suffixes can,

accidentally, produce a form that is homographic with

another word form. For example, the non-vocalized

form "

ﺪﺳﺃ" (>sd) can give the meaning of the verb

ﱡ

ﺪ

ﺳ

ﺃ" (>asud~u) ("

ﱡ

ﺪ

ﺳ +

ﺃ": I block) or the noun "

ﺪ

ﺳ

ﺃ"

(lion; >asaduN). Likewise, clitics can possibly

produce a form that is homographic with another

whole word. For example, the form "

ﻲﺴﻔﻧ" (nfsy) can

be "

ﻲ

ﺴ

ﻔَﻧ" (naf°siy) which means "psychological" or

"myself" which is the combination of "

ﻱ + ﺲﻔﻧ".

3.3 Ambiguity of Non-vocalized Texts

A word is less ambiguous if it is presented in its

vocalized form. Non-vocalized words generate many

solutions for morphological analysis. A form of a

vocalized word can give various morphological

interpretations (Habash and Rambow, 2007; Habash

et al., 2009b) by adding short vowels. For example,

the non-vocalized form "

ﺝﺮﺧﺃ" (>xrj) can accept

about thirty analyses. Among them we cite "

ﺝ

ﺮ

ﺧ

ﺃ" (I

go out; >ax°ruju), "

ﺝ

ﺮ

ﺧ

ﺃ" (he brought out; >ax°raja)

and "

ﺝ

ﺮ

ﺧ

ﺃ" (he got out; >ux°raja). Thus,

orthographic alternation operations often produce

inflected forms which may belong to several different

lemmas or stems. Two words are different just

because one of its middle characters is duplicated

(chedda). For the previous example, "

ﺝ

ﺮ

ﱠ

ﺧ

ﺃ" (I make

out; u>x°riju) is different from the other words only

by adding the double "ﺥ" (xa) character.

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

854

3.4 Discussion

Morphological ambiguities affect other levels of

analysis and mislead IRKES results. Syntactically, it

is difficult to identify the grammatical function of a

word in a sentence. For example, the expression

ﻞﺟﺮﻟﺍ ﺚﺤﺑ" (bHv Alrjl) can be interpreted as "

ﺚ

ﺤ

ﺑ

ﻞ

ﺟ

ﱠ

ﺮﻟﺍ" (baHava Alrajulu) which is a whole sentence

meaning "The man sought" where the first word

ﺚ

ﺤ

ﺑ" (baHava) is the verb of the sentence . It can

also be read as "

ﻞ

ﺟ

ﱠ

ﺮﻟﺍ

ﺚ

ﺤ

ﺑ" (baH°vu Alr~ajuli)

which represents a compound noun meaning "The

search for man".

Likewise, the structure of phrases or expressions

in Arabic can affect morphological disambiguation.

We are mainly talking about a phenomenon

commonly recognized in Arabic texts, known as "free

word order". For example, the previous expression

ﻞﺟﺮﻟﺍ ﺚﺤﺑ" can be replaced by " ﺚﺤﺑ ﻞﺟﺮﻟﺍ" without

changing the meaning. At the semantic level, and

because of the agglutination, the derivation and the

inflection of the Arabic language, the word "

ءﻮﺿﻭ"

(wDw’) can have several meanings depending on its

morphological interpretation. It can be analyzed as

ءﻮ

ﺿ

ﻭ" (ablution; wuDuw’), "

ءﻮ

ﺿ

ﻭ" (water for

ablution; waDuw’) or "

ﻮ

ﺿ" (light; Daw°'uN). In this

example, the letter "

ﻭ" is interpreted as either a

conjunction, or like the first letter of the lemma. Even

in the second case, we get two possible lemmas

diacritized differently.

The example of the sentence "

ﻞﺟﺮﻟﺍ ﺪﻴﺣﻭ ﺐﻫﺫ"

(*hb wHyd Alrjl) matches the syntactic and semantic

level. Indeed, the words of this sentence are all

ambiguous. "

ﺐﻫﺫ" (*Hb) can be "

ﺐ

ﻫ

ﺫ" (gold;

*ahabuN) or "

ﺐ

ﻫ

ﺫ" (he's gone; *ahaba). The word

ﺪﻴﺣﻭ" (WHyd) can be the proper name "

ﺪﻴ

ﺣ

ﻭ"

(Wahid; waHiduN) or "

ﺪﻴ

ﺣ

ﻭ" (alone; waHid) or

ﺪ

ﱠ

ﻴ

ﺣ

ﻭ" (and he neutralized; wa Hay~ada) where the

word is an agglutination of the conjunction "

ﻭ" with

the verb "

ﺪ

ﱠ

ﻴ

ﺣ". The word "ﻞﺟﺮﻟﺍ" can be "

ﻞ

ﺟ

ﱠ

ﺮﻟﺍ"

(man) or "

ﻞ

ﺟ

ﺮﻟﺍ" (leg). The combination of the

solutions of each word gives multiple meanings of

this sentence. The phrase can be interpreted as "

ﺐ

ﻫ

ﺫ

ﻞ

ﺟ

ﺮﻟﺍ

ﺪﻴ

ﺣ

ﻭ" (one-legged man's gold). From these

examples, it is clear that short vowels have great

importance in understanding the POS, function and

meaning of words. Thus, vocalized texts are less

ambiguous than non-vocalized ones. Several

disambiguation approaches have been developed to

overcome the problem of morphological ambiguity.

We detail in the following our main contributions to

deal with the morphological ambiguity of Arabic

texts.

4 THE PROPOSED APPROACH

We present our approach for machine-learning

morphological dependencies to disambiguate Arabic

texts. This approach avoids manual user intervention

in the training phase by exploiting vocalized texts

which are less ambiguous. We model the

disambiguation task as a classification problem as it

is already used in several existing systems (Khoja,

2001; Diab et al., 2004; Habash and Rambow, 2005;

Habash and Rambow, 2007; Roth et al., 2008; Habash

et al., 2009b; Bounhas et al., 2015ab; Ayed, 2017;

Ayed et al., 2018).

We present, in Figure 1, the general architecture

of the proposed approach. We use vocalized Arabic

texts in both training and testing steps. Firstly, ML

classifiers are trained using vocalized texts. Secondly,

we remove short vowels from these texts to perform

the testing step and compute disambiguation rates. A

morphological analyzer gives the different analysis

values of each word. An analysed word corresponds

to the values of the 14 morphological features (POS,

Adjective, Aspect, Case, Conjunction, Determiner,

Gender, Mode, Number, Particle, Person,

Preposition, Voice, Pronoun). Training and testing

datasets are presented in a unified format. Indeed, to

resolve the ambiguity of a morphological feature

(MF), we first define the appropriate attributes that

describe each instance of the training set. These

attributes are also those of the test set.

An attribute, or a morphological characteristic of

a given word, is related to the characteristics of its

preceding and following words. We specify a window

that controls the number of words considered as

attributes describing the class of an instance. In many

existing approaches, the window size is equal to 2

(Habash and Rambow, 2005). Thus, to classify the

of a specific word w, we define the attributes MF-

2, MF-1, MF+1 and MF+2 if the window size is 2

(Ayed et al., 2012b). We can spread the classification

attributes to 56 forming the attributes’ values of MF

± p with i ∈ [| 1,14 |]; where 14 is the number of

attributes and p ∈ {1, 2} (Ayed et al., 2012a). They

indicate, respectively, the morphological attributes of

the two preceding words and the two following words

of w. Thus, we use the other morphological

characteristics of the two neighbouring words to

determine the class value of the current word. The

class value is known in the training set. It corresponds

Experimenting Machine-Learning Algorithms for Morphological Disambiguation of Arabic Texts

855

to the value of the analysis of the MF

attribute of the

word w. However, in the test set, the class is

unknown, and the values of the morphological

attributes can be ambiguous since they correspond to

the analysis of non-vocalized words. This is because

the building of test instances differs from the building

of the training instances only in the input data which

are non-vocalized words. Therefore, the class is the

morphological feature to be disambiguated.

Figure 1: Overview of the approach architecture.

Unlike existing tools (Diab et al., 2004; Habash

and Rambow, 2005; Habash and Rambow, 2007;

Habash et al., 2009b) whose learn from an annotated

corpus, we build our architecture-based ML

classifiers from unannotated texts. This training

method is widely used for unsupervised

disambiguation. In our case, the context used to

resolve the ambiguity of a given word is itself

ambiguous since an attribute in a training set can have

several values which correspond to the possible

analysis of a vocalized word. Indeed, a vocalized

word itself can be ambiguous. This kind of problem

can be considered as a case of imprecision (Bounhas

et al., 2015ab).

ML classifiers used for morphological

disambiguation accept only perfect data with training

and test instances involving complete and exact

information. For this purpose, we propose a method

of transforming imperfect data into perfect data, with

the aim at aligning them with the input format of ML

classifiers. The main idea of our approach is to

transform training and/or test instances containing

imperfect data into data with precise attributes and

classes. We describe, in the following, the operating

mode of the proposed approach to transform data, in

order to overcome the problem of imperfection.

Table 1: Example of imperfect instances of the training set.

POS-1 POS+1 POS (class)

NOUN VERB-PERFECT {NOUN ;

NOUN-PROP}

{NOUN ;

VERB-PERFECT

}

NOUN-PROP NOUN

NOUN-PROP NOUN NOUN

NOUN VERB-PERFECT VERB-PERFECT

We illustrate, through Table 1, an example of an

imperfect training dataset. We assume that the POS is

the morphological feature (MF) to be disambiguated.

We also assume that the attributes are POS-1 and

POS+1. The training set includes 4 instances. The

first instance is uncertain since it provides two

possible values of the class (NOUN and NOUN-

PROP). The second instance is imprecise because it

gives two possible values of the POS-1 attribute

(NOUN and VERB-PERFECT).

We transform this base of instance to obtain a

perfect dataset without loss of information. To solve

the imprecision problem, we denote the values of

attribute A as A

= {a

, a

, ..., a

}. Firstly, we associate

attribute A with each of its a

values to form a new

attribute denoted "A_a

". Secondly, and given that the

POS-1 attribute has 3 possible values in the previous

dataset (see Table 1), we associate 3 new attributes

with it (POS-1_NOUN, POS-1_VERB-PERFECT

and POS-1_NOUN-PROP). Thirdly, we assign

binary values (0 or 1) to the new attributes. For a

given instance, if a

belongs to the values of attribute

A, then the attribute "A_a

" is equal to 1. Finally, we

generate a new training set with precise attributes

given by Table 2. To overcome the problem of class

uncertainty, we propose the decomposition of an

instance into other instances with a single class value.

For an instance with n possible values of a class {c

, ..., c

}, we get n instances associated with it. Each

of these instances has the same attribute values and

class c

. For the example of Table 2, we produce a

training set whose attributes and classes are

completely perfect (cf. Table 3). Thus, this

transformation method overcomes the problem of

imperfect data, and enables us to run ML classifiers

for Arabic texts disambiguation since training and

testing steps require precise and certain instances.

Arabic

vocalized texts

Arabic

non-vocalized

texts

Removing of

short vowels

Morphological

Analysis

(ARAMORPH)

Morphological

features

Imperfect

data

Perfect

data

Disambiguation of

Arabic non-

vocalized texts

ML classifiers

Data transformation

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

856

Table 2: Sample of a transformed training set with precise attributes.

POS-1_ NOUN

POS-1_VERB

PERFECT

POS-1_ NOUN-

PROP

POS+1_NOUN

POS+1_VERB-

PERFECT

POS+1_

NOUN-PROP

POS (class)

1 0 0 0 1 0

{NOUN ;

NOUN-PROP}

1 1 0 0 0 1 NOUN

0 0 1 1 0 0 NOUN

1 0 0 0 1 0 VERB-PERFECT

Table 3: A training set transformed with certain classes.

POS-1_NOUN

POS-1_VERB-

PERFECT

POS-1_

NOUN-PROP

POS+1_

NOUN

POS+1_VERB-

PERFECT

POS+1_

NOUN-PROP

POS (class)

1 0 0 0 1 0 NOUN

1 0 0 0 1 0 NOUN-PROP

1 1 0 0 0 1 NOUN

0 0 1 1 0 0 NOUN

1 0 0 0 1 0 VERB-PERFECT

5 EXPERIMENTATION AND

RESULTS

We present in this section a brief description of the

test collection (cf. Section 5.1), the experimental

scenario and results (cf. Sections 5.2 and 5.3,

respectively). Finally, we discuss a comparative study

highlighting the efficiency of each ML classifiers.

The statistical significance improvements of the best

classifiers are also investigated in Section 5.4.

5.1 Test Collection

The main objective of our approach is to train ML

classifiers (i.e., acquire morphological dependencies)

using vocalized texts, then the testing task is

performed using non-vocalized texts.

 The training

process of the morphological dependencies of the

Hadith Arabic texts has been performed through the

morphological analyzer of vocalized text

"ARAMORPH" to obtain the values of the 14

morphological features. Then, a step of eliminating

short vowels is essential to be able to test on non-

vocalized texts.

Furthermore, we consider the classical Arabic

texts, which have been ignored in previous works.

Therefore, we use a collection of Arabic stories

namely the "Kunuz" corpus of Hadith texts, which

have been studied in several works such as (Harrag et

al., 2009b; Bounhas et al., 2010; Bounhas et al.,

2011b; Harrag et al., 2013; Bounhas et al., 2020). The

Hadiths cover religious knowledge as well as

common and universal knowledge. Moreover, the

Kunuz corpus is one of the rare vocalized Arabic

corpora.

We use the six well-recognized encyclopedic

books organized by theme namely "

ﻱﺭﺎﺨﺒﻟﺍ ﺢﻴﺤﺻ",

(Sahih Al-Bukhari) "

ﻢﻠﺴﻣ ﺢﻴﺤﺻ" (Sahih Muslim),

ﺩﻭﺍﺩ ﻲﺑﺃ ﻦﻨﺳ" (Sunan Abi Dawud), "ﻱﺬﻣﺮﺘﻟﺍ ﻦﻨﺳ"

(Sunan Ettermidhi), "

ﻪﺟﺎﻣ ﻦﺑﺇ ﻦﻨﺳ" (Sunan Ibn

Majah) and "

ﻲﺋﺎﺴ

ﻨﻟﺍ ﻦﻨﺳ" (Sunan Annasaii) (Al-

Echikh, 1998). We limit our experiments to three sub-

corpora corresponding to the following areas of

interest: "

ﺔﺑﺮﺷﻷﺍ" (Al>$RBP; drinks), "ﺝﺍﻭ

ﺰﻟﺍ"

(AlzwAj; marriage) and "

ﺓﺭﺎﻬ

ﻄﻟﺍ" (AlThArp;

purification) (Ayed et al., 2012b).

These areas were chosen because they are generic

and exist in the various books of Hadith. Table 4

presents the number of words in the three sub-corpora

for the six books. While Table 5 summarizes the data

size for the 14 morphological features with their

corresponding total number of attributes and

instances.

Table 4: The number of words in the three sub-corpora for

the six books.

Hadith Boo

Drinks Marria

e Purification Total

Sahih Al-Bukhari 02766 11521 11016 25303

Sahih Musli

09117 06693 05063 20873

Sunan Abi Dawu

02672 05780 15319 23771

Sunan Ettermidhi 01835 05910 09291 17036

Sunan Ibn Majah 01748 06539 13179 21466

Sunan Annasaii 06703 08741 12554 27998

Total 24841 45184 66422 136447

Experimenting Machine-Learning Algorithms for Morphological Disambiguation of Arabic Texts

857

Table 5: Summary of the data size for the 14 morphological

features.

Morphological

feature

Size (Mo) Attributs Instances

POS 215 1961 38304

ADJECTIVE 11.3 105 37562

ASPECT 22.6 209 37567

CASE 22.5 209 37564

CONJUNCTION 11.3 105 37562

DETERMINER 33.7 313 37631

GENDER 16.8 157 37562

MODE 16.9 157 37562

NUMBER 22.4 209 37563

PARTICLE 22.6 209 37713

PERSON 22.5 209 37571

PREPOSITION 11.4 105 37562

VOICE 17.0 157 37562

PRONOUN 358 3329 37615

5.2 Experimental Scenario

We have conducted a set of experimental tests using

7 types of classic classifiers based on 20 machine-

learning algorithms such as: (1) "Bayes classifiers"

include Naïve Bayes and Bayes Net algorithms. (2)

"Function classifiers" are based on SVM (Support

Vector Machine), Logistic and SMO (Sequential

Minimal Optimization) algorithms. (3) "Lazy

classifiers" involve the algorithms IBK (k-nearest

neighbors (k-NN)), KSTAR (K*) and LWL (Locally

Weighted Learning) algorithms. (4) "Meta

classifiers" contain

 ClassificationViaRegression,

FiltredClassifier and Vote algorithm. (5) "MISC

classifiers" incorporate InputMappedClassifier and

SerializedClassifier. (6) "Rules classifiers" are based

on DecisionTable, ZeroR and OneR classifiers. (7)

"Trees classifiers" integrate J48, RandomForest,

RandomTree and DecisionStump algorithms.

We have applied a 10-fold cross-validation

technique (Kohavi, 1995) on the three domains of

application extracted from the six books of Hadith to

estimate the performance of the 20 ML classifiers.

For each morphological feature, we calculate the

average disambiguation rate over the (9 + 1)

iterations. To benefit from these rates, we have

proceeded as follows: (i) we analyze the vocalized

texts and we store the correct morphological

solutions; (ii) we remove short vowels from the same

texts; (iii) we disambiguate the texts obtained with the

20 ML classifiers, then we store the results; and (iv)

we compare the two results to compute the

disambiguation rate.

https://www.cs.waikato.ac.nz/ml/weka/

Firstly, we have experimented these ML

classifiers with their default parameters in WEKA

tool. Then, we have optimized these parameters

(Pedro and Pazzani, 1997) to enhance the

disambiguation rate of each classifier. We have

performed the optimisation process of these ML

classifiers using some WEKA meta-classifiers such

as Grid search, Threshold selector and CVParameter

selection.

We have used in our optimisation process the

CVParameter selection which is the most popular and

efficient. We have fixed the Confidence Factor C

(e.g., C ranged from 0.1 to 0.5 with 5 steps classifiers

options) and modify the Minimum Object M and

vice-versa until achieving the best disambiguation

rate. A comparative study between our default and

optimized results is discussed in Section 5.3. Then an

investigation of the statistical significance

improvement of each optimized classifier is presented

in Section 5.4.

5.3 Experimental Results

We assess the classical ML classifiers in terms of

morphological disambiguation rate. We compare the

results with those given by the known efficient

classifiers SVM and Naïve Bayes for the 14

morphological features. Imperfect test instances

require a transformation process to align with the

input format of classical ML classifiers. These

instances have imperfect attributes and classes. We

illustrate, in Tables 6-9, the morphological

disambiguation rates of the 14 morphological features

given by the above mentioned 20 ML classifiers using

their default and optimal parameters.

The experiments show that the SVM, Naïve Bayes

and Decision Tree classifiers, with their optimal

parameters, have the best average disambiguation

rates of 81.45%, 80.75% and 80.51%, respectively.

For some morphological feature, we obtain the same

results by certain classifiers such as SVM and SMO

having very similar algorithms. This can be explained

by the fact that the associated morphological features

have few (less than 6) possible values. On the other

hand, some other features generate different results

by different classifiers. For example, the attribute

PRONOUN has 64 possible values. The results

highlight that the good optimization of the parameters

of any ML classifier improve its precision when

disambiguating classical Arabic texts.

We note here that some ML classifiers are lacked

by the data size of some morphological features (e.g.,

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

858

POS and PRONOUN), whose cannot be handled even

with the maximum size of WEKA memory (2020

Mo). This problem has been solved by considering

some parts of these data (randomly selected) rather

than working with the full data size. This process

decreases the disambiguation rate of these big-sized

morphological features if compared to others.

5.4 Comparison and Discussion

The goal is to investigate the statistically significant

improvements of the 20 optimized ML classifiers in

terms of disambiguation rate. For this purpose, we use



Wilcoxon Matched-Pairs Signed-Ranks Test

proposed by (Demsar, 2006). It is a non-parametric

alternative to the paired t-test that allows us to

compare two classifiers based on multiple features.

The p-values are calculated by comparing the best

SVM classifier with the 19 other remaining ones (cf.

Table 10). Then, the second-best classifier Naïve

Bayes is also compared to the 18 remaining classifiers

(cf. Table 11). The statistically significant

improvement of a given classifier over a second one

is confirmed if the computed p-value < 0.05.

The p-values of Tables 10 and 11 are less than 0.05

(indicated by *) for the classifiers IBK, KSTAR, LWL,

InputMappedClassifier, Stacking and OneR. We note

here that the difference between the median of the

reference algorithm SVM (respectively Naïve Bayes)

and that of the remaining algorithms is statistically

significant. Whereas the p-values are greater than

0.05 for the other remaining algorithms. In this case,

we do not have sufficient evidence to conclude that

the median of SVM classifier (respectively Naïve

Bayes) is statistically different from the median of the

remaining algorithms.

Table 6: Disambiguation rates of 14 morphological features using default and optimal classifiers’ parameters (1/4).

Morphological

feature

SVM Naïve Bayes Decision Tree Bayesian Net Vote

Default Optimal Default Optimal Default Optimal Default Optimal Default Optimal

POS

89.98 % 92.71% 43.35% 79.54% 71.61% 76.81% 42.97% 52.02% 29.96% 38.25%

ADJECTIVE

96.51% 97.72% 96.51% 97.71% 96.51% 97.72% 96.51% 97.72% 96.51% 97.72%

ASPECT

71.20% 77.56% 71.20% 77.56% 71.20% 77.56% 71.20% 77.56% 71.21% 77.56%

CASE

56.12% 68.21% 56.12% 68.21% 56.12% 68.21% 56.12% 68.21% 56.12% 68.21%

CONJUNCTION

83.03% 87.62% 83.03% 87.62% 83.03% 87.62% 83.03% 87.62% 83.03% 87.62%

DETERMINER

64.12% 67.67% 64.12% 67.67% 64.16% 67.67% 64.12% 67.67% 64.12% 67.67%

GENDER

57.15% 63.77% 57.15% 63.77% 57.15% 63.77 % 57.15% 63.77% 57.15% 63.77%

MODE

99.32 % 99.38% 99.32% 99.38% 99.32% 99.38% 99.32% 99.38% 99.32% 99.38%

NUMBER

85.18% 93.43% 85.18% 93.43% 85.18% 93.43% 85.18% 93.43% 85.18% 93.43%

PARTICLE

96.65% 98.81% 96.65% 98.81% 96.65% 98.81% 96.65% 98.81% 96.65% 98.81%

PERSON

60.22% 67.55% 60.22 % 67.55% 60.22% 67.55% 60.22% 67.55% 60.22% 67.55%

PREPOSITION

82.87% 85.12% 82.87% 85.12% 82.87% 85.12% 82.87% 85.12% 82.87% 85.12%

VOICE

71.21% 77.92% 71.21% 77.92% 71.21% 77.92% 71.21% 77.92% 71.21% 77.92%

PRONOUN

56.88 % 62.81% 59.06 % 66.24% 62.38% 65.61% 61.08% 67.70% 58.94% 63.33%

Average

76.46% 81.45% 73.29% 80.75% 75.54% 80.51% 73.40% 78.89% 72.32% 77.60%

Table 7: Disambiguation rates of 14 morphological features using default and optimal classifiers’ parameters (2/4).

Morphological

feature

SMO J48 RandomForest DecisionStump Logistic

Default Optimal Default Optimal Default Optimal Default Optimal Default Optimal

POS

43.02% 52.85% 58.64% 64.68% 29.96% 34.58% 29.96% 33.61% 27.09% 35.12%

ADJECTIVE

96.51% 97.72% 96.51% 97.72% 96.51% 96.58% 96.51% 96.58% 96.51% 96.58%

ASPECT

71.20% 77.56% 71.20% 77.56% 71.20% 77.27% 71.20% 77.27% 71.20% 77.27%

CASE

71.20% 68.21% 56.12% 68.21% 56.12% 68.30% 56.12% 68.30% 56.12% 68.30%

CONJUNCTION

83.03% 87.62% 83.03% 87.62% 83.03% 86.60% 83.03% 86.60% 83.03% 86.60%

DETERMINER

64.16% 67.67% 64.16% 67.67% 64.16% 65.64% 64.12% 68.81% 64.12% 68.81%

GENDER

57.15% 63.77% 57.15% 63.77% 57.15% 65.72% 57.15% 65.72% 57.15% 65.72%

MODE

99.32% 99.38% 99.32% 99.38% 99.32% 99.40% 99.32% 99.40% 99.32% 99.40%

NUMBER

85.18% 93.43% 85.18% 93.43% 85.18% 88.59% 85.18% 88.59% 85.18% 88.59%

PARTICLE

96.65% 98.81% 96.65% 98.81% 96.65% 96.80% 96.65% 96.80% 96.65% 96.80%

PERSON

60.22% 67.55% 60.22% 67.55% 60.22% 69.82% 60.22% 69.82% 60.22% 69.82%

PREPOSITION

82.87% 85.12% 82.87% 85.12% 82.87% 85.71% 82.87% 85.71% 82.87% 85.71%

VOICE

71.21% 77.92% 71.21% 77.92% 71.21% 80.02% 71.21% 80.02% 71.21% 80.02%

PRONOUN

61.58% 66.67% 61.58% 66.67% 62.39% 64.55% 60.93% 64.05% 61.58% 64.52%

Average

74.52% 78.88% 74.56% 79.72% 72.57% 77.11% 72.46% 77.23% 72.30% 77.38%

Experimenting Machine-Learning Algorithms for Morphological Disambiguation of Arabic Texts

859

Table 8: Disambiguation rates of 14 morphological features using default and optimal classifiers’ parameters (3/4).

Morphological

feature

FilteredClassifier DecisionTable Classification

ViaRegression

ZeroR IBK (KNN)

Default Optimal Default Optimal Default Optimal Default Optimal Default Optimal

POS

55.79% 60.79% 48.68% 52.29% 43.10% 50.10% 29.96% 37.19% 61.36% 62.36%

ADJECTIVE

96.51% 96.58% 96.51% 96.58% 96.51% 96.58% 96.51% 96.58% 96.51% 96.55%

ASPECT

71.20% 77.27% 71.20% 77.27% 71.20% 77.27% 71.20% 77.27% 71.20% 73.20%

CASE

56.12% 68.30% 56.12% 68.30% 56.12% 68.30% 56.12% 68.30% 56.12% 57.12%

CONJUNCTION

83.03% 86.60% 83.03% 86.60% 83.03% 86.60% 83.03% 86.60% 83.03% 83.50%

DETERMINER

64.12% 68.81% 64.12% 68.81% 64.12% 68.81% 64.12% 68.81% 64.18% 65.18%

GENDER

57.15% 65.72% 57.15% 65.72% 57.15% 65.72% 57.15% 65.72% 57.15% 58.15%

MODE

99.32% 99.40% 99.32% 99.40% 99.32% 99.40% 99.32% 99.40% 99.32% 99.33%

NUMBER

85.18% 88.59% 85.18% 88.59% 85.18% 88.59% 85.18% 88.59% 85.18% 87.18%

PARTICLE

96.65% 96.80% 96.65% 96.80% 96.65% 96.80% 96.65% 96.80% 96.65% 96.65%

PERSON

60.22% 69.82% 60.22% 69.82% 60.22% 69.82% 60.22% 69.82% 60.22% 65.22%

PREPOSITION

82.87% 85.71% 82.87% 85.71% 82.87% 85.71% 82.87% 85.71% 82.87% 84.87%

VOICE

71.21% 80.02% 71.21% 80.02% 71.21% 80.02% 71.21% 80.02% 71.21% 73.21%

PRONOUN

61.75% 64.27 % 62.30% 65.51% 58.68% 63.12% 58.94% 62.41% 61.67% 63.67%

Average

74.37% 79.19% 73.90% 78.67% 73.24% 78.35% 72.32% 77.37% 74.76% 76.16%

Table 9: Disambiguation rates of 14 morphological features using default and optimal classifiers’ parameters (4/4).

Morphological

feature

KSTAR LWL InputMapped

Classifier

Stacking OneR

Default Optimal Default Optimal Default Optimal Default Optimal Default Optimal

POS

62.02% 63.02% 35.53% 36.53% 29.97% 30.97% 29.97% 30.97% 32.69% 33.69%

ADJECTIVE

96.51% 96.55% 96.51% 96.55% 96.51% 96.55% 96.51% 96.55% 96.51% 96.55%

ASPECT

71.20% 73.20% 71.20% 73.20% 71.20% 73.20% 71.20% 73.20% 71.20% 73.20%

CASE

56.12% 57.12% 56.12% 57.12% 56.12% 57.12% 56.12% 57.12% 56.12% 57.12%

CONJUNCTION

83.03% 83.50% 83.03% 83.50% 83.03% 83.50% 83.03% 83.50% 83.03% 83.50%

DETERMINER

64.12% 65.12% 64.12% 65.12% 64.12% 65.12% 64.12% 65.12% 64.12% 65.12%

GENDER

57.15% 58.15% 57.15% 58.15% 57.15% 58.15% 57.15% 58.15% 57.15% 58.15%

MODE

99.32% 99.33% 99.32% 99.33% 99.32% 99.33% 99.32% 99.33% 99.32% 99.33%

NUMBER

85.18% 87.18% 85.18% 87.18% 85.18% 87.18% 85.18% 87.18% 85.18% 87.18%

PARTICLE

96.65% 96.65% 96.65% 96.65% 96.65% 96.65% 96.65% 96.65% 96.65% 96.65%

PERSON

60.22% 65.22% 60.22% 65.22% 60.22% 65.22% 60.22% 65.22% 60.22% 65.22%

PREPOSITION

82.87% 84.87% 82.87% 84.87% 82.87 % 84.87% 82.87% 84.87% 82.87% 84.87%

VOICE

71.21% 73.21% 71.21% 73.21% 71.21% 73.21% 71.21% 73.21% 71.21% 73.21%

PRONOUN

58.58% 60.58% 61.69% 63.69% 58.55% 60.55% 58.55% 60.55% 60.47% 63.47%

Average

74.58% 75.98% 72.91% 74.31% 72.29% 73.69% 72.29% 73.69% 72.62% 74.09%

Table 10: The p-values for the statistically significant improvements of SVM compared to the 19 classifiers.

SVM

vs.

Naïve

Bayes

Decision

Tree

Bayesian

Net

Vote SMO J48

Random

Forest

Decision

stump

Logistic

Filtred

Classifier

0.59 0.65 0.65 0.65 0.65 0.65 0.63 0.94 0.94 0.94

Decision

Table

Classification

ViaRegression

ZeroR

IBK

(KNN)

KSTAR LWL

InputMapped

Classifier

Stacking OneR

0.89 0.94 0.75 0.001* 0.0009* 0.001* 0.0009* 0.0009* 0.001*

Table 11: The p-values for the statistically significant improvements of Naïve Bayes compared to the 18 classifiers.

Naïve

Bayes

vs.

Decision

Tree

Bayesian

Net

Vote SMO J48

Random

Forest

Decision

stump

Logistic

Filtred

Classifier

Decision

Table

0.28 1 0.28 1 1 0.36 0.59 0.68 0.63 0.72

Classification

ViaRegression

ZeroR

IBK

(KNN)

KSTAR LWL

InputMapped

Classifier

Stacking OneR

0.55 0.55 0.0009* 0.0009* 0.0009* 0.0009* 0.0009* 0.0009*

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

860

6 CONCLUSION

We have focused in this paper on the problem of

morphological disambiguation of Arabic texts. The

latter have a very rich and complex morphology that

has given rise to many challenges for the natural

language processing tasks. The morphological

disambiguation process still among the main

challenge for the information retrieval systems. It is

useful to identify the appropriate form of index terms

in Arabic IR. In this context, we have proposed a new

architecture for morphological disambiguation of

Arabic terms using a series of classical ML

algorithms. During the data pre-processing step, we

have proposed and implemented a data

transformation technique useful to transform

imperfect data to perfect ones. Then, the selected

classifiers are trained on vocalized Arabic texts and

tested on non-vocalized ones. We have also

performed an optimisation process to enhance the

efficiency of each classifier,

a method that none has

suggested before to improve the morphological

disambiguation rate of Arabic texts.

We have experimented these ML classifiers using

the "Kunuz" collection of classical Arabic texts in

order to compare and discuss their efficiency. The

SVM classifier seems to be the most efficient in the

morphological disambiguation of Arabic texts. It

achieved a statistically significant improvement over

a few competing algorithms. Besides, the second

efficient Naïve Bayes classifier has achieved some

statistically significant improvements compared to

some ML algorithms. Our short-term concern is to

use Friedman's statistical test to compare the 20 ML

classifiers together to more investigate the degree of

statistically significant improvement of each

algorithm. But our long-term concern consists in

testing these ML classifiers using the modern Arabic

texts collection TreeBank.

ACKNOWLEDGEMENTS

This work was funded by the Liwa College of

Technology in Abu Dhabi (UAE) under research

grant IRG-BIT-002-2020.

REFERENCES

Al-Ansary, S. (2005). Building a Computational Lexicon

for Arabic: A corpus-based approach. In Alhawary, M.

T., Benmamoun, E., (eds.): Current Issues in Linguistic

Theory, volume 267, pp. 173–193. John Benjamins

Publishing Company, Amsterdam.

Al-Echikh, A. (1998). Encyclopedia of the six major

citation collections. Dar-esselem, Ryadh, KSA.

Ayed, R. (2017). Désambiguïsation Morphologique de

Textes Arabes à Base de Classification Possibiliste

pour la Recherche d'Information Socio-Sémantique.

PhD Thesis, ENSI, Manouba University, Tunisia.

Ayed, R., Bounhas, I., Elayeb, B., Evrard, F., Bellamine

Ben Saoud, N. (2012a). Arabic Morphological Analysis

and Disambiguation Using a Possibilistic Classifier. In

Proceedings of ICIC-2012, pp. 274–279, Huangshan,

China. Springer Berlin Heidelberg.

Ayed, R., Bounhas, I., Elayeb, B., Evrard, F., Bellamine

Ben Saoud, N. (2012b). A Possibilistic Approach for

the Automatic Morphological Disambiguation of

Arabic Texts. In Proceedings of SNPD-2012, pp. 187–

194, Kyoto, Japan, IEEE Computer Society.

Ayed, R., Chouigui, A., Elayeb, B. (2018a). A New

Morphological Annotation Tool for Arabic Texts. In

Proceedings of AICCSA-2018, pp. 1-6, Aqaba, Jordan,

IEEE Computer Society.

Ayed, R., Elayeb, B., Bellamine Ben Saoud, N. (2018b).

Possibilistic Morphological Disambiguation of

Structured Hadiths Arabic Texts Using Semantic

Knowledge. In Proceedings of ICAART-2018, pp. 565-

572, Funchal, Madeira, Portugal, SciTePress,.

Belguith, L. H., Chaâben, N. (2006). Analyse et

désambiguïsation morphologiques de textes arabes non

voyellés. In Proceedings of TALN-2006, pp. 493–501,

Leuven, Belgique, ATALA.

Ben Khiroun, O., Ayed, R., Elayeb, B., Bounhas, I.,

Bellamine Ben Saoud, N., Evrard, F. (2014). Towards

a New Standard Arabic Test Collection for Mono- and

Cross-Language Information Retrieval. In Proceedings

of NLDB-2014, LNCS 8455, pp. 168–171, Montpellier,

France, Springer International Publishing.

Bounhas, I., Ayed, R., Elayeb, B., Bellamine Ben Saoud, N.

(2015b). A hybrid possibilistic approach for Arabic full

morphological disambiguation. Data & Knowledge

Engineering, 100:240-254.

Bounhas, I., Ayed, R., Elayeb, B., Evrard, F., Bellamine

Ben Saoud, N. (2015a). Experimenting a discriminative

possibilistic classifier with reweighting model for

Arabic morphological disambiguation. Computer

Speech & Language, 33(1):67-87.

Bounhas, I., Elayeb, B., Evrard, F., Slimani, Y. (2010).

Toward a computer study of the reliability of Arabic

stories. Journal of the American Society for

Information Science and Technology, 61(8):1686–

1705.

Bounhas, I., Elayeb, B., Evrard, F., Slimani, Y. (2011b).

Organizing Contextual Knowledge for Arabic Text

Disambiguation and Terminology Extraction.

Knowledge Organization, 38(6):473–490.

Bounhas, I., Soudani, N., Slimani, Y. (2020). Building a

morpho-semantic knowledge graph for Arabic

information retrieval. Information Processing &

Management, 57(6): 102124

Experimenting Machine-Learning Algorithms for Morphological Disambiguation of Arabic Texts

861

Bousmaha, K. Z., Rahmouni, M. K., Kouninef, B.,

Belguith, L. H. (2016). A Hybrid Approach for the

Morpho-Lexical Disambiguation of Arabic. Journal of

Information Processing Systems, 12(3):358–380.

Daoud, D. (2009). Synchronized Morphological and

Syntactic Disambiguation for Arabic. Advances in

Computational Linguistics, 41:73–86.

Daoud, D., Daoud, M. (2009). Arabic Disambiguation

Using Dependency Grammar. In Proceedings of TALN-

2009, Senlis, France, ATALA.

Demsar, J. (2006). Statistical Comparisons of Classifiers

over Multiple Data Sets. Journal of Machine Learning

Research, 7:1–30.

Diab, M., Hacioglu, K., Jurafsky, D. (2004). Automatic

Tagging of Arabic Text: From Raw Text to Base Phrase

Chunks. In Proceedings of HLT-NAACL 2004: Short

Papers, HLT-NAACL-Short’04, pp. 149–152,

Stroudsburg, PA, USA. Association for Computational

Linguistics.

Elayeb, B. (2009). SARIPOD : Système multi-Agent de

Recherche Intelligente POssibiliste des Documents

Web. PhD. Thesis in Artificial Intelligence, INPT,

Toulouse, France, 2009.

Elayeb, B. (2018). Recherche d'Information Possibiliste :

De la Désambiguïsation et la Reformulation de

Requêtes vers la Fiabilité de l'Information Recherchée.

HDR Thesis, ENSI, Manouba University, Tunisia.

Elayeb, B. (2019). Arabic Word Sense Disambiguation: A

Review. Artificial Intelligence Review, 52(4):2475-

2532.

Elayeb, B. (2021). Arabic Text Classification: A Literature

Review. In Proceedings of AICCSA-2021, pp. 1-8,

Tangier, Morocco, IEEE Computer Society.

Elayeb, B., Bounhas, I. (2016). Arabic Cross-Language

Information Retrieval: A Review. ACM Transactions

on Asian and Low-Resource Language Information

Processing, 15(3):18:1-18:44.

Elayeb, B., Chouigui, A., Bounhas, M., Ben Khiroun, O.

(2020). Automatic Arabic Text Summarization Using

Analogical Proportions. Cognitive Computation, 12(5):

1043-1069.

Elayeb, B., Evrard, F., Zaghdoud, M., Ben Ahmed, M.

(2009). Towards an intelligent possibilistic web

information retrieval using multiagent system. The

International Journal of Interactive Technology and

Smart Education, 6(1):40–59.

ElHadj, Y., Al-Sughayeir, I. A., Al-Ansari, A. M. (2009).

Arabic part-of-speech tagging using the sentence

structure. In Proceedings of the 2

International

Conference on Arabic Language Resources and Tools,

pp. 241–245, Cairo, Egypt.

Habash, N. (2007). Arabic Morphological Representations

for Machine Translation. In Soudi, A., Bosch, A. v. d.,

Neumann, G., (eds.): Arabic Computational

Morphology, Vol. 38, Speech and Language

Technology, pages 263–285. Springer Netherlands.

Habash, N., Rambow, O. (2005). Arabic Tokenization,

Part-of-speech Tagging and Morphological

Disambiguation in One Fell Swoop. In Proceedings of

ACL’05, pp. 573–580, Stroudsburg, PA, USA.

Association for Computational Linguistics.

Habash, N., Rambow, O. (2007). Arabic Diacritization

Through Full Morphological Tagging. In Proceedings

of HLT-NAACL 2007: Short Papers, HLT-NAACL-

Short’07, pp. 53–56, Stroudsburg, PA, USA.

Association for Computational Linguistics.

Habash, N., Rambow, O., Roth, R. (2009b).

MADA+TOKAN: A toolkit for Arabic tokenization,

diacritization, morphological disambiguation, POS

tagging, stemming and lemmatization. In Proceedings

of the 2

International Conference on Arabic

Language Resources and Tools (MEDAR), pp. 102–

109, Cairo, Egypt.

Harrag, F., Alothaim, A., Abanmy, A., Alomaigan, F.,

Alsalehi, S. (2013). Ontology Extraction Approach for

Prophetic Narration (Hadith) using Association Rules.

International Journal on Islamic Applications in

Computer Science and Technology, 1(2).

Harrag, F., Hamdi-Cherif, A., Malik, A., Al-Salman, S., El-

Qawasmeh, E. (2009b). Experiments in improvement

of Arabic information retrieval. In proceedings of the

International Conference on Arabic Language

Processing (CITALA), Rabat, Morocco.

Khoja, S. (2001). APT: Arabic part-of-speech tagger. In

Proceedings of the NAACL-2001, Pennsylvania, USA.

Kohavi, R. (1995). A Study of Cross-validation and

Bootstrap for Accuracy Estimation and Model

Selection. In Proceedings of IJCAI’95, Volume 2, pp.

1137–1143, San Francisco, CA, USA. Morgan

Kaufmann Publishers Inc.

Mansour, S., Sima’an, K., Winter, Y. (2007). Smoothing a

Lexicon-based POS Tagger for Arabic and Hebrew. In

Proceedings of the Workshop Semitic’07, pp. 97–103,

Stroudsburg, PA, USA. Association for Computational

Linguistics.

Othman, E., Shaalan, K., Rafea, A. (2004). Towards

resolving ambiguity in understanding Arabic sentence.

In Proceedings of the International Conference on

Arabic Language Resources and Tools (NEMLAR), pp.

118–122, Cairo, Egypt.

Pedro, D., Pazzani, M. (1997). On the optimality of the

simple Bayesian classifier under zero-one

loss. Machine Learning, 29:103–137.

Roth, R., Rambow, O., Habash, N., Diab, M., Rudin, C.

(2008). Arabic Morphological Tagging, Diacritization,

and Lemmatization Using Lexeme Models and Feature

Ranking. In Proceedings of HLT-NAACL-Short’08,

HLT-Short’08, pp. 117–120, Stroudsburg, PA, USA.

Association for Computational Linguistics.

Tlili-Guiassa, Y. (2006). Hybrid Method for Tagging

Arabic Text. Journal of Computer Science, 2(3):245–

248.

Zribi, C. B. O., Torjmen, A., Ahmed, M. B. (2006). An

Efficient Multiagent System Combining POS-Taggers

for Arabic Texts. In Proceedings of CICLing-2006,

LNCS 3878, pp. 121–131, Mexico City, Mexico,

Springer Berlin Heidelberg.

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

862