Experimenting Machine-Learning Algorithms for Morphological

Disambiguation of Arabic Texts

Bilel Elayeb

1,2 a

, Mohamed Firas Ettih

3b

and Raja Ayed

2,4 c

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

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

potential of words’ vowels in a text and interprets

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

Numerous applications in the field of Arabic NLP

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

phonetization and automatic Arabic text

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 "

interpreted as a noun meaning "food" (

phenomenon represents a challenge for

morphologically rich languages such as Arabic (Diab

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

ensuring that documents and queries are represented

orthographic variations including syntactic and

richness causes ambiguity and difficulty in

type of texts, namely modern Arabic texts. POS

disambiguation consists of determining the

context. One of the important steps in the

of Arabic texts using several morphological

test

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

Arabic texts. This corpus contains Hadith texts

assigned to the Prophet of Islam Mohammad

follows: existing works on morphological

disambiguation of Arabic texts are summarized and

morphological ambiguity are presented in Section 3.

Section 4 presents the architecture of the proposed

approach as well as the training and testing data

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

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

POS tagger system that combines morphological

analysis and the Markov model. The tagging process

authors have relied on the online morphological

measure of over 80%.

Bounhas et al. (2015a) have proposed three

disambiguation: (i) the first classifier is based on

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

the combination of these two measures. The authors

hybrid possibilistic approach that combines the

possibilistic classifier with linguistic rules to assign

approach has improved the disambiguation rate of

Arabic texts. The authors have also presented an

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

taken advantage of the XML format of the structured

involved semantic attributes in their possibilistic

improvement in the disambiguation rates of

knowledge.

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-

English. The Arabic word "

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

analyzer must segment this word to recognize the

or nouns’ declension generate inflection of the words

forms. The word "

of the verb

inflectional operations produce a slight change in

pronunciation without an explicit effect on spelling

voice, "

voice and "

eat".

The Arabic word is the result of a combination of

root, vowels, prefixes, infixes, suffixes and a

produce a form that is homographic with another

whole word. For example, the form "

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 "

about thirty analyses. Among them we cite "

"

(Wahid; waHiduN) or "

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

deal with the morphological ambiguity of Arabic

4 THE PROPOSED APPROACH

We present our approach for machine-learning

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

we remove short vowels from these texts to perform

the testing step and compute disambiguation rates. A

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.

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

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

attributes to 56 forming the attributes’ values of MF

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

of w. Thus, we use the other morphological

unknown, and the values of the morphological

the analysis of non-vocalized words. This is because

of the training instances only in the input data which

morphological feature to be disambiguated.

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

several values which correspond to the possible

information. For this purpose, we propose a method

of transforming imperfect data into perfect data, with

classifiers. The main idea of our approach is to

transform training and/or test instances containing

classes. We describe, in the following, the operating

mode of the proposed approach to transform data, in

order to overcome the problem of imperfection.

NOUN VERB-PERFECT VERB-PERFECT

We illustrate, through Table 1, an example of an

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-

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

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

"

(Sunan Ettermidhi), "

Majah) and "

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

features.

DETERMINER 33.7 313 37631

MODE 16.9 157 37562

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

 ClassificationViaRegression,

application extracted from the six books of Hadith to

estimate the performance of the 20 ML classifiers.

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

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

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

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

We have used in our optimisation process the

CVParameter selection which is the most popular and

in Section 5.4.

morphological disambiguation rate. We compare the

results with those given by the known efficient

morphological features. Imperfect test instances

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.