Possibilistic Morphological Disambiguation of Structured Hadiths

Arabic Texts using Semantic Knowledge

Raja Ayed

1,2

, Bilel Elayeb

1,3

and Narjès Bellamine Ben Saoud

RIADI Research Laboratory, ENSI, Manouba University, 2010, Tunisia

Faculty of Sciences of Gabes, Gabes University, 6072, Tunisia

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

Keywords: Possibility Theory, Arabic Morphological Disambiguation, Hadiths Structured Corpus, Classification

Attributes.

Abstract: We propose, in this paper, a possibilistic morphological approach to disambiguate hadiths Arabic texts

using semantic knowledge. The disambiguation is considered as a classification problem. The possibilistic

approach uses vocalized texts to train a possibilistic classifier in order to classify non-vocalized texts as they

are more ambiguous. Morphological attributes are used for training and test. Hadiths are structured in XML

format that provides semantic information. We enlarge the classification attributes’ set by adding semantic

attributes extracted from the hadiths structure. We prove that the possibilistic approach gives the best rates

using AlKhalil analyzer to prepare the training and the test sets. Our proposed possibilistic approach

enhances disambiguation rates of Arabic hadiths’ texts when it includes semantic knowledge.

1 INTRODUCTION

Arabic belongs to the Semitic languages and is

considered as the major language of the Middle East.

It is written from right to left. It has many special

properties and a complex method of words’

construction especially for the classical Arabic.

Morphological disambiguation approaches deal with

the complex morphology of Arabic language to

resolve ambiguity. Vocalized texts are considered

less ambiguous than non-vocalized ones. Hadiths

corpus presents an interesting data source of

classical and vocalized texts that can be used to

evaluate a morphological disambiguation approach.

As the disambiguation is considered as a

classification task, we propose the morphological

possibilistic approach that learns morphological

knowledge from vocalized texts and exploit this

knowledge to disambiguate non-vocalized ones. To

do so, we organize this paper as follows: we give, in

section 2 and section 3, an overview of the Arabic

morphological ambiguity and the morphological

disambiguation approaches. In section 4, we

illustrate the possibilistic disambiguation approach.

Section 5 describes the hadiths corpus and focuses

on the socio-semantic aspects of hadiths. In section

6, we describe our approach that takes advantage of

the hadiths structure and their semantic knowledge

to disambiguate morphological features using the

possibilistic disambiguation approach. Subsequently,

a set of experimentations results are presented in

section 7. Finally, we summarize in section 8 our

findings in the conclusion and propose future

perspectives.

2 ARABIC MORPHOLOGICAL

AMBIGUITY

The morphological analysis tries to identify possible

solutions, of a given word, of some morphological

features like POS, gender, aspect, case, etc. Some

morphological analysis tools give the stem (Hajic,

2000) or the lemma (Attia, 2008) of a word to

determine correctly its morphological features. They

give, also, a segmented solution specifying the

prefixes, affixes and suffixes. A word is considered

ambiguous if it has more than just one solution.

Disambiguation assigns, to the ambiguous word, the

most appropriate analysis given the context of the

word (Ayed et al., 2012). Disambiguation becomes a

challenging field of many NLP researches (Ayed et

al., 2012). This issue is closely related to the

Ayed, R., Elayeb, B. and Saoud, N.

Possibilistic Morphological Disambiguation of Structured Hadiths Arabic Texts using Semantic Knowledge.

DOI: 10.5220/0006651105650572

In Proceedings of the 10th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2018) - Volume 2, pages 565-572

ISBN: 978-989-758-275-2

565

complexity of the morphology of Arabic language.

Ambiguity is mainly caused by the agglutination

and the lack of diacritics (short vowels). So, the

same word could have many interpretations (Habash

et al., 2009). For example, the word “” is

agglutinated. It means in English “Will you

remember us?”. In this case, morphological analyzer

aims to remove the prefixes and suffixes to extract

the root of the word and give the right

morphological feature. Some prefixes and suffixes

can be homographic with each other (Attia, 2008).

Without diacritics, the word “” (qrAt) can mean

“















“ (qaraAtu; I read) or “















” (qaraAta; you

read). Some other meanings can be identified by

adding more short vowels, or diacritics. Prefixes and

suffixes can by coincidence, produce a form that is

homographic with another full form. For instance,

the vocalized word “















” (>ak°ramu) can be, at the

same time, (i) the proper noun “Akram” in some

grammatical cases and (ii) the adjective “more

generous”.

Morphological ambiguities can reach even the

syntactic and the semantic levels of analysis. At the

syntactic level, the sentence may give different

grammatical functions if it contains one or more

ambiguous words (Bounhas et al., 2011). For

example, all the words composing the sentence "

" (*hb wHyd Alrjl) are ambiguous. ””

can be “











” (*ahabu; gold) or “











” (*ahaba; he

goes). “” can be the proper noun “















“

(waHiydN; Wahid) or “















” (waHiydu; single).

“” can be “











” (Alrrajulu; the man) or “











”

(Alrrijli; the leg). The combination of the solutions,

of each word, may give multiple meanings of this

sentence. This affects, even, the semantic level of

ambiguity. The sentence may be interpreted as, for

example, “



































" (the gold of the man with

a single leg). These illustrations show that diacritics,

or shorts vowels, are essential in determining the

grammatical category, the morphological feature and

the significance of some words. Therefore, vocalized

texts are considered less ambiguous than non-

vocalized texts.

3 ARABIC MORPHOLOGICAL

DISAMBIGUATION

Researches, in Arabic NLP, are interested in

disambiguation of Arabic words by identifying their

morphological features depending on their context.

Several classification methods are used to

disambiguate 1morphological words’ features

(Habash et al., 2009; Roth et al., 2008). Hence, the

disambiguation is considered as a classification

problem where a morphological feature of an

ambiguous word is the class that we aim to identify.

The classification methods are based on learning

techniques to train classifier from the datasets where

their morphological features are known.

The existing approaches in disambiguation can

be divided into three categories that are: (i) linguistic

approaches, (ii) statistical approaches and (iii)

hybrid approaches.

Linguistic approaches are, also, called rule-based

approaches. They use rules written by linguists to

tag the different morphological features (Diab et al.,

2007). We talk about heuristics, contextual and non-

contextual rules (Elshafei et al., 2002).

Statistical approaches incorporate different

classification methods such as Support Vector

Machine (SVM) and Hidden Markov models

(HMM) to calculate probability of each value of a

grammatical category of a word (Vapnik, 1999).

SVM was used by the disambiguation tools MADA

(Habash et al., 2013) and MADAMIRA (Pasha et

al., 2014) that combines MADA and AMIRA

(Habash et al., 2009). To define the grammatical

category of a word, MADAMIRA converts the

Arabic text to the Buckwalter (2004) transliteration.

Then, it analyzes the text morphologically and gives

all the possible POS for each word. Finally, it

applies SVM and language models to choose the

right POS from all proposed POS values after the

morphological analysis.

The hybrid approaches combine the two last

approaches. We talk about linguistic and statistical

approaches to disambiguate the words in their

morphological level. The approach of Tlili-Guiassa

(2006) is based on the MBL approach (Memory

based learning). It analyzes grammatical and

inflectional affixes and grammatical rules. This

approach is used to classify a collection of Quranic

and educational texts. (Zribi et al., 2006) combine

rule-based approach with a trigram HMM tagger.

Texts with 6000 words are used to train the trigram

classifier and heuristic rules were applied to select

from the proposed results.

According to (Hoceini et al., 2011), the linguistic

approaches for disambiguation are more reliable

than the statistical approaches. The linguistic

approaches, which are based on a specific set of

rules, require only a linguist to define these rules.

While for the statistical approaches, the statistics

calculated for training are the same used for any test

domain. The training phase is necessary for both

statistical and hybrid approaches in order to learn the

required settings for disambiguation.

ICAART 2018 - 10th International Conference on Agents and Artiﬁcial Intelligence

566

The hybrid approaches combine the linguistic

and statistical approaches so they take advantage of

each approach. As a result, they are considered as

the most effective and coherent in terms of analysis.

4 THE POSSIBILISTIC

MOPHOLOGICAL

DISAMBIGUATION

APPROACH

We present the possibilistic disambiguation

approach based on the possibility theory (Zadeh,

1978). This approach was proposed by Bounhas et

al. (2015). The possibilistic theory is unlike the

probability theory. It discerns between uncertainty

and imprecision which describe an incomplete

information (Ayed et al., 2012).

Imprecision is revealed when a reality state is

defined by variables with multi-values. Uncertainty

appears when we are unable to provide or to know a

statement to determine the real value of a proposal

(Dubois and Prade, 2010).

The main idea of the possibilistic approach is to

provide a possibilistic classifier (Haouari et al.,

2009) that obtains disambiguation knowledge from

vocalized texts and tests on non-vocalized texts.

Training and test phases involve instances that use

classification attributes and classes. Each instance is

associated to a word w. For the possibilistic

disambiguation approach, classes are the 14

morphological features (MF) presented by MADA

(Roth et al., 2008). They are, exclusively, POS,

conjunction, particle, determiner, pronoun, person,

voice, aspect, gender, number, case, preposition,

mode and adjective (Ayed et al., 2012). Each

morphological feature may have one or more

possible values. The training and the test attributes

designate the morphological features (MF) of the

two following and the two preceding words of w

(MF±i, i{1, 2}). For instance, we get the attributes

POS-2, POS-1, POS+1 and POS+2 which designate

the POS (Part-Of-Speech) of the two previous and

following words of w. POS may be the class of w.

Every instance, in the training set, is related to a

vocalized word. Table 1 gives an example of a

training set where the training attributes are (POS±i,

i=1) and the class is POS. Table 2 presents an

example of a testing set that corresponds to a non-

vocalized word whose POS is ambiguous and needs

to be determined. The attributes of the test set must

be the same of the training set.

In table 1 and table 2, both training and test sets

are imperfect since they contain uncertain and

imprecise instances. In fact, the second instance, of

the training set, is uncertain as it gives more than

one class’ value. The first instance, of the training

set, is imprecise since it contains attribute that

provide more than just one value; i.e. POS+1. The

test instance is also imprecise. Indeed, its POS+1

attribute has two possible values. The possibility

theory deals with the problem of imperfection of

training and test instances. Arabic vocalized words

are less ambiguous but, in some cases, they give

ambiguous values which justify the imperfection of

particular instances of the training sets.

Disambiguation consists in identifying the right

values, of the morphological features (POS,

conjunction, case, etc.), among values given by a

morphological analyzer.

Table 1: Example of instances from a training dataset.

Word

POS-1

POS+1

POS (class)















VERB_

PERFECT

{INTERROG_

PART; REL_

PRON}

VERB_

PERFECT











VERB_

PERFECT

VERB_

PERFECT

{NOUN;

NOUN-

PROP}

Table 2: Example of an instance from a test dataset.

Word

POS-1

POS+1

POS (class)



VERB_

PERFECT

{INTERROG_

PART; REL_

PRON}

4.1 The Training Phase

To prepare training datasets, we analyze vocalized

texts and rearrange instances as shown in table 1.

Morphological analysis tools identify the different

features of a given word out of its context (Hajic,

2000). Possibilistic disambiguation approach uses

the updated version of BAMA 1.2.1 (Buckwalter,

2004), AraMorph that treats vocalized texts. For

each AraMorph identifies the prefixes, the suffixes

and the stem of each word and associates values of

the 14 morphological features (Habash et al., 2009;

Roth et al., 2008). Possibilistic disambiguation uses

a possibilistic classifier where the class MF is one of

the morphological features (MF {POS,

conjunction, particle, determiner, pronoun, person,

voice, aspect, gender, number, case, preposition,

mode and adjective}) and the attributes are the set of

(MF±i, i{1, 2}). We denote c

a value of a class MF

and a

a value of an attribute MF±i (that we note

). In the training phase, we compute the frequency

Possibilistic Morphological Disambiguation of Structured Hadiths Arabic Texts using Semantic Knowledge

567

measure of an attribute value (a

) for a class c

(Elayeb et al., 2009). We determine this measure for

each couple (attribute value, class) in the training

set. The frequency is computed as follows:









































(1)

Where 







 is the number of instances that

have the class 



and 



values for the attribute A





 is the number of possible values of A

. The Max

operator is used to normalize the frequency (Elayeb

et al., 2009).



















  











(2)

T is the training set, |T| is the number of instances in

the training set, I

is an instance from T. |A

| is the

number of the attribute A

values in I

and |C

| is the

class values of I

. The product |A

|*|C

| decodes the

imperfection of an instance. In fact, if the instance I

is imprecise than one or more of its attributes have

more than one value. Consequently, |A

| is not equal

to 1 for the imprecise attribute A

and |C

| is not

equal to 1 for the uncertain class. Hence, if the

instance is perfect, than the product is equal to 1.



ijkL

is equal to 1 if the value a

belongs to the

possible values of A

in the instance I

and c

is one of

the possible classes of I

. If not,



ijkL

is equal to 0.

4.2 The Test Phase

To prepare test datasets, we analyze non-vocalized

texts and rearrange instances as shown in table 2.

Morphological disambiguation of a non-vocalized

word consists in determining the accurate values of

its morphological features. In other terms,

morphological disambiguation determines the class

value of the test instance.

We reduce the ambiguity of some Arabic words

by including linguistic rules defined in (Diab et al.,

2007). Linguistic rules decrease the number of the

test instances. We present each rule as a function of

the attributes used by the possibilistic classifier. For

instance, we experiment the rule “A preposition

cannot follow a preposition” by: if (POS-1=”PREP”)

then POS≠”PREP”. Hence, we remove the “PREP”

value from the possible POS values. For the

remaining ambiguous test instances, we use the

possibilistic morphological disambiguation approach

to select the morphological feature value.

The possibilistic morphological disambiguation

approach proposes measures to distinguish the

accurate class; i.e. necessity and possibility

(Bounhas et al., 2015). Each of these measures is

computed, over the training set, for each possible

value of the class, to determine among the accurate

value. A morphological feature’s value is accorded,

to a non-vocalized word, if it has the maximum

value of possibility and necessity.

The possibility and the necessity measures are

given, respectively, by the formulas (3) and (4).

















 







  

















(3)

















    

  





























(4)

We denote the test instance I

. m is the number of

test attributes. c

is a class value. A

is a test attribute

and |A

| is the number of the attribute A

values in the

instance I

. a

is a value of A

. The frequencies









are previously calculated over the

training set.  is equal to log

(P/nc

) where P is

the number of possible class values and nc

is the

number of the class values having a non-null

frequency with the attribute a

i. e. Freq (a

, c

)≠0.





is a factor that we add for imprecise attributes. In

fact, if an attribute has 4 possible values, we

compute the product of the frequencies of these 4

values and we multiply (for the possibility) or divide

(for the necessity) each of these frequencies by 1/4

(



. The selected class value of the test instance I

corresponds to the value c*. It is the value that has

the highest score among all class values:



























 















(5)

The possibilistic approach is considered as a hybrid

approach as it combines possibilistic measures with

linguistic rules.

5 THE HADITHS ARABIC

CORPUS

The hadiths Arabic corpus

represents Islamic texts,

said by the prophet Mohamed, which discusses

several real-life concerns. Hadiths are written in

classical Arabic with vocalized words. This corpus

was the topic of some research fields as knowledge

extraction, texts classification (Harrag et al., 2009;

Harrag et al., 2013) and information retrieval

(Bounhas et al., 2010; Bounhas et al., 2011). Ben

ICAART 2018 - 10th International Conference on Agents and Artiﬁcial Intelligence

568

Khiroun et al. (2014) attempt to create, from the

hadiths texts, a standard test collection for Arabic

information retrieval.

The corpus of hadiths is one from the rare

vocalized Arabic corpora. It contains about 2.5

million of words dispersed on more than 100 books

of hadiths. The most known and used books are

Sahih Muslim, Sahih Al Bukhari, Sunan Annasaii,

Sunan Ibn Majah, Sunan Abi Dawud, and Sunan

Ettermidhi (Al-Echikh, 1998). Besides, the corpus is

entirely structured into chapters and sub-chapters

which provide an appropriate source of contextual

data (Bounhas et al., 2011). Hadiths corpus can be

studied along several social and semantic axes which

make it a reliable resource for information retrieval

and knowledge extraction.

5.1 Social Aspect in Hadiths

Each hadith is composed of two parts (i)

the Sanad () or the chain of narrators through

which the hadith was transmitted and (ii) the Matn

() or the text told by the prophet. An example of

a hadith is given by figure 1.

Mussad told us from Yahia from Shu’bah from

Qatadah from Anas that the prophet said: None of you

will be a true believer until he loves for his brother that

which he loves for himself. "Al Bukhari recited it ".













































































































































































































































































”

Figure 1: An example of a vocalized hadith.

The Sanad is represented by 













































































































































and the Matn gives

the replied text 







































































































.

The narrators are separated by the word   (En ;

from). The verb  (qAl ; said) announces the

beginning of the text transmission. The expression

   (Al Bukhari recited it) means that the

hadith exists in "Sahih Al Bukhari". The sheikh

(;

$yx) is an Islamic scholar. The sheikh Al Bukhari

learned the hadith text from his sheikh Musadd (











)

who heard it from his sheikh Yahya (











). The latter

learned the text from his sheikh Shu’bah (











) who

heard it from Qatadah (











) who was a disciple of

one of the prophet companions Anas (







). Anas

transmitted the hadith text from the prophet.

Social relationships can be extracted from the

Matn and the Sanad. We can identify the sheikh

relationship commonly through the word . Also,

family relationships can be identified, for instance,

by the words  (<bn; son of) or  (>bw; father

of). A narrator can also report what his brother said

using, for example, 



 (my brother told me)

or what his grandfather said using, for example,









 (my grandfather told me). The narrators in

the Sanad can be related through other social

attributes such as place of residence and

communities’ membership.

5.2 Semantic Aspect in Hadiths

The hadiths’ books are structured by theme, except

for some books that are arranged by narrators.

Descriptive data of several hadiths were added by

the scholars to facilitate the understanding of the

texts.

The hadiths are structured in XML format in

such a way that they are classified according to

specific themes. A hadith can cover a multitude of

themes and a huge amount of information in the

Matn as well as in the Sanad. From a hadith, we can

extract information from the knowledge conveyed

by the content or by the titles of the chapters and the

sub-chapters and even the comments provided by the

narrators and scholars of the hadith. The XML tags

add semantic information to the hadiths. In figure 2,

we present an example of a structured XML file

from Al Bukhari book. The <THEME> tag indicates

the theme. <S> designates the section where the

hadith appears. The <S> element is followed by all

its hadiths tags. <DOC> describes a hadith where

<R> indicates its Sanad and <TEXT> includes its

Matn. The hadith’s theme is that of its section.

…

<THEME> 































 </ THEME>

…

</S>

<DOC>

…

<R ID="4698" S="



" TP="F">





































































</R>

…

<TEXT> 















































































































 















 















 











 















































 

























































































 "

</TEXT>

</DOC>

Figure 2: An example of a structured XML hadith.

Possibilistic Morphological Disambiguation of Structured Hadiths Arabic Texts using Semantic Knowledge

569

6 POSSIBILISTIC

MORPHOLOGICAL

DISAMBIGUATION OF

HADITHS

Many research works in information retrieval are

carried out on hadiths texts (Harrag et al., 2009; Jozi

et al., 2012). Morphological disambiguation is

crucial to get pertinent results of an information

retrieval system (Bessou and Touahria, 2014). It’s

used to disambiguate both documents and queries

which are written, commonly, in non-vocalized

texts.

We take advantage of the hadiths structure and

their provided semantic data to disambiguate a non-

vocalized hadith text using the possibilistic

approach. To do, we perform a morphological

analysis of the hadiths texts (an example is given in

figure 2) using a morphological analyzer to

determine the various values of the 14

morphological features. The analysis gives the

different possibilities without affecting the exact

value of a morphological attribute.

We add other classification attributes to provide

more sense to morphological data. Classification

attributes include:

▪ A semantic attribute that designates the theme of

the analyzed word. In fact, if the word belongs to

a hadith text, than its theme is that of this hadith.

In other terms, a word that appears in a <DOC>

element gets the theme of the hadith presented by

this <DOC> element.

▪ An attribute, that we denote “dimension”, which

indicates whether a word belongs to a Sanad or a

Matn. In fact, if a word appears in the <R>

element content, than it belongs to Sanad and if a

word appears in the <TEXT> element content,

than it belongs to Matn.

▪ The morphological attributes (i.e. MF±i, i{1,

2}) used by the possibilistic morphological

disambiguation approach.

Training and test instances use the same

classification attributes. Table 3 shows an example

of an ambiguous non-vocalized word. To simplify

the example, we use only the POS-1 and POS+1

morphological attributes. The non-vocalized word

 (<bn) is ambiguous even if we add short vowels.

In fact, the vocalized word 







 (<ib°ni) can mean (i)

“son of” which is a noun and (ii) “build” which is a

verb. Thus, the possible POS values set is

{VERB_IMPERATIVE; NOUN}. The theme given

to this word is  (Al<ymAn; the faith). This

word is a part of Sanad. In a narrative context, the

word  cannot be an imperative verb. Based on the

necessity and the possibility measures computed

over the training set, the class NOUN gives the

maximum sum value.

Table 3: Example of a test instance using morphological

and semantic attributes.

Word

POS-1

POS+1

theme

Dim-

ension

POS

(class)



NOUN_

PROP

NOUN_

PROP



Sanad

7 EXPERIMENTAL RESULTS

We describe, in this section, the evaluation method

used to experiment the possibilistic disambiguation

approach. We present, also, disambiguation results

of hadiths using morphological and semantic

information. We study variation effect of the

analyzers on the disambiguation results.

7.1 The Evaluation Method

To assess the performance of the possibilistic

approach on the hadiths disambiguation, we use the

5-fold cross-validation. In other terms, 80% of the

corpus texts are used for training and 20% are used

for test. We obtain five possible combinations. The

training texts are vocalized. We omit vowels of the

test set. Hence, we obtain non-vocalized texts. We

compute the average success rates from all the 4+1

combinations. To get these rates, we analyze the

vocalized texts and we save their morphological

solutions. Then, we omit the short vowels of the

same texts. Finally, we apply the possibilistic

classifier to disambiguate these texts and we save

the results. If the two obtained results of a word are

similar, then this word is correctly classified.

7.2 Comparing Morphological

Analyzers for Possibilistic

Disambiguation

We propose to vary several morphological analyzers

in order to determine the best one for the

possibilistic disambiguation approach. We

disambiguate about 10000 words from hadiths texts

using (i) Aramorph, (ii) BAMA analyzer of

MADAMIRA (Pasha et al., 2014) and (iii) AlKhalil

analyzer (Bousmaha et al., 2013). Table 4 illustrates

the results of the possibilistic disambiguation

approach using these analyzers. The results are

given for the POS morphological feature. The

ICAART 2018 - 10th International Conference on Agents and Artiﬁcial Intelligence

570

disambiguation based on AlKhalil gives the highest

rate followed by the MADAMIRA analyzer.

Aramorph gives a respectable disambiguation rate

but it doesn’t overcome Alkhalil and MADAMIRA.

Hence, the morphological disambiguation approach

depends, closely, on the analyzer used to prepare

training and test instances.

Table 4: Disambiguation rates using different

morphological analyzers.

Analyzer

Aramorph

MADAMIRA

AlKhalil

disambiguation

rate

93.93%

94.09%

94.67%

7.3 Evaluating the Possibilistic

Disambiguation Approach

Morphological disambiguation consists in

determining the accurate values of the

morphological features. We perform experiments on

the book Al Bukhari using AlKhalil analyzer to

provide classification instances. The disambiguation

rates of hadiths’ texts give 76.36%, 76.46% and

76.23% respectively using Decision Tree, SVM and

Naïve Bayesian classifiers (Bounhas et al., 2015).

Comparing to these rates, our approach using the

possibilistic classifier gives better results. In fact, the

possibilistic morphological disambiguation approach

provides an average rate of 86.37% for the 14

morphological features using only the morphological

attributes MA (i.e. MF±i, i{1, 2}) for

classification. The average increases by 0.29% if we

use both the morphological attributes and the

semantic attributes (SA) in classification (see section

6). We notice that improvement does not affect all

the morphological features. Indeed, only the

disambiguation rates of POS, aspect, adjective,

conjunction, gender, person, preposition and

pronoun increase. This is explained by the fact that

adding semantic attributes, for these features,

reduces the training set size. The possibility and the

necessity measures computing over the training

instances give high results of the correct classes. For

instance, for the POS class, some values of

morphological attributes, such as NOUN_PROP

(proper noun), appear most in narrators chains

(Sanad). Thus, an ambiguous word that contains

NOUN_PROP in its possible values of POS is

probably a proper noun rather than another value of

the POS class. Similarly, many Arabic proper nouns

can have the same meaning as an adjective. We note,

for example, “” (>sEd; more happy) or “”

(jmyl; beautiful). This fact may justify the

enhancement of the disambiguation rate of the

adjective feature using semantic attributes. Thus, the

addition of semantic attributes reduces the number

of possible values in a particular context. Semantic

attributes restrict and filter the training sets which

increase the possibility and necessity measures

giving support to the accurate class.

Table 5: Possibilistic morphological disambiguation rates

using morphological attributes and semantic attributes.

Disambiguation

rates (MA)

Disambiguation

rates (MA & SA)

POS

95.13%

96.01%

ADJECTIVE

99.17 %

99.27%

ASPECT

81.53 %

81.58%

CASE

63.52 %

63.52%

CONJUNCTION

91.07 %

91.12%

DETERMINER

97.02 %

97.02%

GENDER

96.55 %

96.66%

MODE

99.96%

NUMBER

93.10 %

93.10%

PARTICLE

98.88 %

98.88%

PERSON

66.06 %

67.02%

PREPOSITION

88.27 %

90.02%

VOICE

79.11 %

79.11%

PRONOUN

59.81 %

59.83%

Average

86.37 %

86.66%

8 CONCLUSION

We presented, in this paper, the possibilistic

morphological disambiguation approach that uses

morphological and semantic attributes to

disambiguate 14 morphological features of hadiths

texts. Disambiguation consists in choosing the

accurate morphological feature value from the

solutions proposed by a morphological analyzer. We

proved that the possibilistic approach gave the best

rates using AlKhalil analyzer. The possibilistic

disambiguation rates that use morphological and

semantic attributes are better than those that use only

morphological attributes. Experiments showed an

encouraging improvement of the possibilistic

approach to deal with classical Arabic texts. We

aim, as a future work, to improve the performance of

an information retrieval system by presenting a

queries and documents disambiguation phase. We

consider that our work is a first step toward building

an information retrieval system which treats both

vocalized and non-vocalized documents and focuses

on hadiths Arabic texts.

Possibilistic Morphological Disambiguation of Structured Hadiths Arabic Texts using Semantic Knowledge

571

REFERENCES

Attia, M., 2008. Handling Arabic Morphological and

Syntactic Ambiguity Within the LFG Framework with

View to Machine Translation. PhD Thesis, University

of Manchester.

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

Ben Saoud, N., 2012b. Arabic Morphological Analysis

and Disambiguation Using a Possibilistic Classifier.

In: Proceedings of the 8

International Conference on

Intelligent Computing (ICIC), July 25-29, 2012 (pp.

274–279). Springer Berlin Heidelberg.

Bessou, S. and Touahria, M., 2014. An Accuracy-

Enhanced Stemming Algorithm for Arabic

Information Retrieval. Neural Network World Journal,

24(2):117–128.

Bounhas, I., Elayeb, B., Evrard, F., Slimani, Y., 2011.

Organizing Contextual Knowledge for Arabic Text

Disambiguation and Terminology Extraction.

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

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

N., 2015. A Hybrid Possibilistic Approach for Arabic

Full Morphological Disambiguation. Data &

Knowledge Engineering, vol. 100,Part B, pp. 240-254.

Bousmaha, K. Z., Abdoun, S. C., Belguith, L. H. et

Rahmouni, M. K., 2013. Une approche de

désambiguïsation morpho_lexicale évaluée sur

l’analyseur morphologique Alkhalil.

www.webreview.dz, 20(2): 32–46.

Buckwalter, T., 2004. Buckwalter Arabic Morphological

Analyzer Version 2.0. Online Publication, Linguistic

Data Consortium (LDC) catalog number

LDC2002L49, ISBN 1-58563-257-0, from

http://www.nongnu.org/aramorph/

Diab, M., Hacioglu, K., Jurafsky, D., 2007. Automated

Methods for Processing Arabic Text: From

Tokenization to Base Phrase Chunking. In van den

Bosch, A. and Soudi, A., editors, Arabic

Computational Morphology: Knowledge-based and

Empirical Methods. Kluwer/Springer.

Dubois, D., Prade, H., 2010. Formal Representations of

Uncertainty. In D. Bouyssou, D. Dubois, M. Pirlot, &

H. Prade (Eds.), Decision-making Process (pp. 85–

156). ISTE &Hoboken, London, UK, Wiley, USA.

Elshafei, M., Al-Muhtaseb, H., Al-Ghamdi, M., 2002.

Techniques for high quality Arabic speech synthesis.

Information Sciences 140(3), 255-267.

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

2009. Towards an intelligent possibilistic web

information retrieval using multiagent system.

Interactive Technology and Smart Education, 6(1),

2009, 40–59.

Habash, N., Rambow, O., Roth, R., 2009.

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),

Cairo, Egypt, April 22-23, pp. 102-109.

Habash, N., Roth, R., Rambow, O., Esk, R. et Tomeh, N.

(2013). Morphological Analysis and Disambiguation

for Dialectal Arabic. In Proceedings of the 2013

Conference of the North American Chapter of the

Association for Computational Linguistics: Human

Language Technologies(NAACL-HLT), Atlanta, GA.

Hajic, J., 2000. Morphological Tagging: Data vs.

Dictionaries. In: Proceedings of the 1

North

American Chapter of the Association for

Computational Linguistics Conference (pp. 94–101).

Stroudsburg, PA, USA: Association for Computational

Linguistics, April 29- May 24, 2000.

Haouari, B., Ben Amor, N., Elouedi, Z., Mellouli, K.,

2009. Naïve possibilistic network classifiers. Fuzzy

Sets and Systems, 160(22), 3224–3238.

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

and El-Qawasmeh, E., 2009. Experiments in

improvement of Arabic information retrieval. In

proceedings of the 3rd International Conference on

Arabic Language Processing, Rabat, Morocco.

Hoceini, Y., Cheragui, M. A., Abbas, M., 2011. Towards a

New Approach for Disambiguation in NLP by

Multiple Criterian Decision-Aid. The Prague Bulletin

of Mathematical Linguistics 95, 19-32.

Jozi, H., Zadeh, A. R., Barati, E. et Minaei, B, 2012. A

new framework for detecting similar texts in Islamic

Hadith Corpora. In The International Conference on

Language Resources and Evaluation, Istanbul,

Turkey.

Pasha, A., Al-Badrashiny, M., Diab, M., El Kholy, A.,

Eskander, R., Habash, N., Pooleery, M., Rambow, O.

and Roth, R. (2014). MADAMIRA: A Fast,

Comprehensive Tool for Morphological Analysis and

Disambiguation of Arabic. In Proceedings of the Ninth

International Conference on Language Resources and

Evaluation (LREC’14), Reykjavik, Iceland.

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 the Association for

Computational Linguistics conference (ACL),

Columbus, Ohio, USA, June 19-20, 2008

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

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

248.

Vapnik, V. N., 1999. An overview of statistical learning

theory. IEEE Transactions on Neural Networks, 10(5),

988–999. DOI:10.1109/72.788640

Zadeh, L.A., 1978. Fuzzy Sets as a Basis for a Theory of

Possibility. Fuzzy Sets and Systems, 1, 3-28, 1978.

Zribi, C., Torjmen, A., Ben Ahmed, M., 2006. An

Efficient Multi-agent System Combining POS-

Taggers for Arabic Texts. In Proceedings of 7

international conference of Computational Linguistics

and Intelligent Text Processing, LNCS Volume 3878,

Springer, 121-131.

ICAART 2018 - 10th International Conference on Agents and Artiﬁcial Intelligence

572