Morphological Disambiguation of Texts

Based on Analogical Proportions

Bilel Elayeb

1,2 a

, Myriam Bounhas

and Mohamed Firas Ettih

RIADI Research Laboratory, ENSI, Manouba University, Tunisia

LARODEC Research Laboratory, ISG Tunis, Tunis University, Tunisia

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

Keywords: Morphological Disambiguation, Analogical Proportions, Machine Learning Algorithms,

Deep Learning Algorithms, Feature Selection, Classification.

Abstract: The Arabic language is known for its complexity, which encompasses extensive morphological and

orthographic variations, as well as significant syntactic and semantic diversity. These unique characteristics

often result in morphological ambiguity in Arabic. In this paper, we tackle the challenge of morphological

disambiguation in Arabic texts. We frame this task as a classification problem, where the possible values of

morphological features represent the classes, and a classification algorithm is used to assign the appropriate

class to each word based on its context. Specifically, we investigate the effectiveness of an analogy-based

classifier for morphological disambiguation in Arabic texts. Analogical Proportions (AP) are statements that

express the relationship between four elements A, B, C, and D such that "A differs from B as C differs from

D". Leveraging Analogical Proportions-based inference, the AP classifier predicts the fourth, unknown

element (D), given that the first three (A, B, and C) are known. We evaluate this analogical classifier using a

corpus of Classical Arabic texts. The average disambiguation rate (74.80%) of the AP classifier outperforms

that of a set of well-established machine-learning and deep learning-based classifiers.

1 INTRODUCTION

Morphological disambiguation in the Arabic

language involves analyzing and clarifying

ambiguities in the structure and meaning of words by

determining their accurate morphological attributes,

such as root, stem, part of speech, and grammatical

features (e.g., gender, number, and tense). Given

Arabic's complex morphology, this process is

essential for natural language processing (NLP) tasks

(Elayeb et al., 2009; Elayeb and Ben Khiroun, 2023),

including machine translation, text-to-speech

systems, and information retrieval.

Arabic Morphological Disambiguation (AMD)

still faces several challenges, such as: (i) rich

morphology: Arabic words often include prefixes,

suffixes, and infixes that convey various grammatical

details, (ii) Complex word structures: words can

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

https://orcid.org/0000-0001-8320-6722

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

exhibit intricate combinations of affixes, clitics, and

roots, which result in more difficult analysis, (iii)

absence of vowels: written Arabic, particularly

Modern Standard Arabic, frequently omits diacritical

marks, which complicate the identification of a

word's correct form, and (iv) ambiguity: the context-

sensitive nature of Arabic leads to multiple possible

morphological interpretations for many words.

Morphological disambiguation poses a significant

challenge for languages rich in morphology and

ambiguity, such as Arabic. A non-vocalized Arabic

word can have more than 12 possible interpretations

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

example, the non-vocalized Arabic word “ﺏﺮﺿ” can

be interpreted as a verb meaning “to strike” (

ﺏ

ﺮ

ﺿ) or

as a noun meaning “a strike” (

ﺏ

ﺮ

ﺿ).

Some Arabic words are homographs, meaning

they are written identically but have different

Elayeb, B., Bounhas, M. and Ettih, M. F.

Morphological Disambiguation of Texts Based on Analogical Proportions.

DOI: 10.5220/0013308300003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 1207-1214

ISBN: 978-989-758-737-5; ISSN: 2184-433X

1207

meanings. Processing these words depends on

recognizing morphemes, which form the basic units

of meaning.

The purpose of morphological analysis is to

determine the various criteria or functions of each

lexical unit or word. Examples of these criteria

include Part-Of-Speech (POS), which identifies

whether a word is a noun, verb, or particle, as well as

features such as number, gender, and information

about clitics. Ambiguity arises during the analysis,

especially when contextual clues do not align with a

word's intended interpretation.

This paper explores a method for addressing

morphological ambiguity in Arabic texts. We

approach this task as a classification problem, where

the possible morphological feature values represent

the classes, and a classification algorithm determines

the correct class for each word based on its context.

Specifically, we investigate the efficiency of an

analogy-based classifier for AMD. This type of

classifier has recently demonstrated its effectiveness

in classifying Arabic texts (Bounhas et al., 2024).

We begin by introducing Analogical Proportions

(AP), which establish a relationship between four

elements, A, B, C, and D, such that "A is to B as C is

to D". Then, based on the analogical inference, the AP

classifier proposed by Bounhas et al. (2017), that we

use, can predict the fourth, unknown element D when

the first three elements (A, B, and C) are provided.

This algorithm has not yet been tested before in the

context of Arabic text morphological disambiguation.

We evaluate the performance of this analogical

classifier on a corpus of Classical Arabic texts and

compare it to both ML and DL classifiers.

The remainder of this paper is organized as

follows: Section 2 provides a summary and

discussion of existing works on morphological

disambiguation in Arabic texts. Section 3 summarizes

a background on analogical proportions. In Section 4,

we describe the proposed analogy-based classifier for

AMD. Section 5 presents the experimental results and

includes a comparative analysis of the AP classifier

with both ML-based and DL-based classifiers.

Finally, Section 6 concludes the study and suggests

directions for future research.

2 RELATED WORKS

Many studies have reduced ambiguity by identifying

the POS feature of Arabic text. POS attribute

disambiguation is the process of determining the

grammatical class of a word in a particular context.

Morphological disambiguation is known as a

classification problem in which a set of POS values

represents a class, and a single classification method

is used to assign each word occurrence to a class

based on sentence context.

One of the most important steps in disambiguation

is choosing an appropriate classification method.

Several automatic classification methods have been

used, leveraging ML techniques to train classifiers

from sentences annotated with POS values. In

literature, disambiguation methods are commonly

organized into three categories: (i) rule-based

methods, (ii) statistical methods, and (iii) hybrid

methods that combine these two approaches, and (iv)

ML and DL-based approaches.

Rule-based or linguistic methods use a set of rules

created by linguists to assign labels to different

morphological attributes. Several WSD systems have

been described by Daoud (2009), mainly involving

heuristic, textual, and non-textual rules. Besides,

these rules typically fall into grammatical, structural,

and logical classes. Daoud and Daoud (2009)

proposed a specialized parser called "EnConverters,"

developed in UNL and a rule-based programming

language. The Daouds defined several types of

disambiguation rules that merge syntactic and

morphological context associations.

Statistical methods build one or more learning

models from annotated corpora. These methods often

utilize statistical models such as the Hidden Markov

Model, which assumes a Markov process with

unknown parameters. Classification methods such as

SVM calculate the probabilities for each grammatical

class of a word. Moreover, MADA is a tool designed

by Habash and Rambow (2007) and is widely used to

resolve morphological ambiguities in Arabic texts.

Then, MADAMIRA (Pasha et al., 2014) combines

MADA and the Arabic tokenizer AMIRA (Diab,

2007).

A hybrid method combines statistical information

with linguistic rules to address morphological

ambiguity. Khoja (2001) is among the researchers

who adopted this approach, implementing it with the

Viterbi algorithm. She calculated two probabilities

from an annotated corpus of 50,000 words: (i)

contextual probability, the likelihood that one label

precedes or follows another, and (ii) lexical

probability, the likelihood that a word has a certain

morphological attribute. Based on these statistics, a

set of grammatical rules was developed, achieving an

accuracy of over 90% (Khoja, 2001).

Belguith and Chaâben (2006) also proposed a

method for morphological analysis and

disambiguation, categorized as a statistical approach

that incorporates rule-based elements. This method

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1208

involves five steps (Ayed et al., 2018b): (i) text

segmentation into words, (ii) morphological

preprocessing by removing clitics based on a

predefined list, (iii) affixal analysis, distinguishing

the root and affixes of each word, (iv) morphological

analysis using MORPH2, and (v) post-processing to

group words using lexicons and a rule set. This

method calculates the morphological attributes of

each word using a term dictionary.

Bousmaha et al. (2016) proposed a hybrid

disambiguation approach focusing on the selection of

diacritical marks across different analysis levels. This

method combines multi-criteria decision-making

with a linguistic approach and offers an alternative

solution for morphological ambiguity. For

evaluation, the authors used an online shape analyzer,

achieving promising results with F-measures above

0.8. Bounhas et al. (2015a) proposed three

possibilistic classifiers for AMD: (i) the first classifier

relies on the possibility measure, (ii) the second on

the necessity measure, and (iii) the third combines

these two measures. They enhanced these classifiers

with information gain scores, serving as weights for

classification attributes, to optimize space

requirements for resolving contextual ambiguity and

thereby simplify the disambiguation process.

Later, Bounhas et al. (2015b) introduced a hybrid

possibilistic approach that integrates the possibilistic

classifier with linguistic rules to assign labels to

various morphological attributes, which improved

Arabic text disambiguation rates. They also addressed

“out-of-vocabulary” words lacking known

morphological analysis. These possibilistic and

hybrid classifiers were evaluated on the Arabic

"Kunuz"

dataset and compared with three machine-

learning (ML) classifiers: SVM, Naïve Bayes, and

Decision Tree.

Besides, Ayed et al. (2018b) explored possibilistic

morphological disambiguation for structured Hadith

texts in Arabic, incorporating semantic knowledge.

Using AlKhalil analyzer, they conducted training and

testing of morphological attributes, leveraging the

XML format of "Kunuz" Hadith texts to integrate

available semantic information. By including

semantic attributes in their possibilistic classifiers,

they achieved improved disambiguation rates in their

experiments.

More recently, Elayeb et al. (2022) experimented

with a range of ML algorithms to address the challenge

of morphological disambiguation in Arabic texts using

various morphological features. The authors evaluated

these algorithms on the Kunuz test collection (Ben

https://github.com/bounhasibrahim/Kunuz/tree/Kunuz-IR

Khiroun et al., 2012a; 2014; Ayed et al., 2018ab),

which comprises classical, vocalized Arabic texts,

specifically Hadith attributed to the Prophet

Muhammad (PBUH). This corpus has attracted

considerable research interest due to its linguistic,

semantic, and social depth, as well as its structured

organization. By contrast, the TREC collections (2001

and 2002), which include Arabic newspaper articles,

lack vowel markers in their texts, contributing to

potential word sense ambiguities and difficulties in

determining POS tags and syntactic roles.

Moreover, Analogical Proportions, first

developed by Prade et al. (2010), have demonstrated

their efficiency when applied in several domains,

such as information retrieval (IR) (Bounhas and

Elayeb, 2019), Arabic text summarization (Elayeb et

al., 2020), Arabic text classification (Elayeb et al.,

2023; Bounhas et al., 2024), as well as the

classification of structured data (Bounhas et al., 2017;

Bounhas and Prade, 2023; 2024). This paper aims to

investigate the performance of such proportions in

disambiguating Arabic text data collections.

3 ANALOGICAL PROPORTIONS

An analogical proportion (AP) is a relationship

between 4 items A, B, C, D ∈ X. This relationship

states that “A differs from B as C differs from D”

which enables us to compare the pair (A, B) to the pair

(C, D) (in terms of similarities and differences). An

Analogical proportion is usually denoted A : B :: C :

D. This proportion also holds when “A : C :: B : D”

(by central permutation), “A : B :: A : B” (by

reflexivity) and “C : D :: A : B” (by symmetry) (Prade

and Richard, 2010).

Boolean Setting: In the Boolean context, there

are only six valid valuations where “A : B :: C : D”

holds true or “A : B :: C : D = 1”. These six valid

valuations, or six possible assignments, are the only

configurations among the 16 possible configurations

that make an analogical proportion true. In particular,

a 4-tuple (A, B, C, D) is in analogical proportion if it’s

in one of those particulars’ assignments: (0,0,0,0)

(1,1,1,1) (0,0,1,1) (1,1,0,0) (0,1,0,1) (1,0,1,0). As can

be seen, the proportion remains valid for these six

patterns even when items are negatively coded.

Nominal Extension: In the nominal case, the

analogical proportion can only be true in only three

possible assignments out of the six. Otherwise, it is

Morphological Disambiguation of Texts Based on Analogical Proportions

1209

no longer true. A : B :: C : D is valid if it is of the form:

(r, r, r, r) or (r, s, r, s) or (r, r, s, s) such that: r ≠ s.

Multiple-Valued Extension: To handle numerical

attributes the domain has to be in the interval [0,1],

which means that the analogical proportion A : B :: C

: D linking the 4 items A, B, C and D takes values in

the interval [0,1] (Dubois et al., 2016). It is then a

matter of obtaining a high or low AP value depending

on whether the values are closer to 1 or 0. For

examples:

• (0, 0.2, 0, 1)

 we expect that A : B :: C : D has a

low value close to 0 since 0.2 is closer to 0.

• (0, 0.9, 0, 1)

 we expect that A : B :: C : D has a

high value close to 1 since 0.9 is closer to 1.

Normalization: Numerical attributes need to be

normalized when dealing with AP classifiers since it

helps reduce the time required to classify, enhances

the classifier’s performance and standardizes the data,

which improves the overall process.

Inference: As introduced above, analogical

proportions have been recently formalized within

Boolean, nominal, and numerical frameworks.

In these latter, given a valid analogical proportion

A : B :: C : D, the inference principle helps to derive

one component of the four-part proportion from the

other three. More formally, if the four objects A, B, C,

D build a valid AP and if the first three objects A, B,

C are known then it is possible to compute the fourth

object D by solving the analogical equation: finding

a value X such that A : B :: C : X = 1.

Since the AP can be valid only for six possible

assignments of the 4-tuples, there are cases where the

equation, in the Boolean case, A : B :: C : X = 1 has

no solution. Indeed, the equations 1 : 0 :: 0 : X = 1

and 0 : 1 :: 1 : X = 1 have no solution.

It has been proven that the above analogical

equation is solvable if and only if (A≡ B)

∨

(A≡ C)

holds. In that case, the unique solution X is X = A ≡

(B ≡ C); thus X is either equal to B (if A = C) or X is

equal to C (if A = B). In this paper, we mainly focus

on the Boolean setting since, as seen later in the

experimental study, the datasets we tested contain

only Boolean features.

4 ANALOGY-BASED

CLASSIFIER FOR AMD

Analogical inference coincides with AP classification

in which the class for an object D can be predicted

based on the known classes of the three others: A, B

and C. The classification of D (unknown variable) is

only possible if the equation on the class, Class(A) :

Class(B) :: Class(C) : X, is solvable. (Bounhas et al.,

2017). In classification problems, we assume that

items are no longer defined as simple variables but

rather as vectors of n attribute values, i.e.

𝐴

⃗

=(a

,…,a

) where

𝑎

𝑖

is the value of attribute

𝑖

for

item 𝐴

⃗

, similarly 𝐵



⃗

=(b

,…,b

), 𝐶

⃗

=(c

,…,c

) and

𝐷



⃗

=(d

,…,d

). We also assume implicitly that the four

items 𝐴

⃗

,𝐵



⃗

,𝐶

⃗

and 𝐷



⃗

are represented in terms of the

same set of attributes. Then an AP: 𝐴

⃗

:𝐵



⃗

∷𝐶

⃗

∶𝐷



⃗

valid if and only if ∀ 𝑖 ∈



1, 𝑛



,𝑎



:𝑏



∷𝑐



:𝑑



Analogical classifiers, which are essentially based

on the above analogical inference, operate by

identifying triplets of examples (𝐴

⃗

,𝐵



⃗

,𝐶

⃗

) in the

training set that form an AP with the item to be

classified (𝐷



⃗

), on all or a maximum number of

features. These classifiers also ensure that the

corresponding analogical proportion equation for the

class has a valid solution.

In its basic formulation, the analogical classifier

(Bounhas et al., 2017) applies this principle to

determine a solution for the class of 𝐷



⃗

that we denote

𝐶𝑙𝑎𝑠𝑠𝐷



⃗

. When the analogical equation for the

attributes holds, it increments the corresponding score

by 1 and assigns the class label with the highest score

to 𝐷



⃗

. This classifier systematically explores all

possible triplets in the training set.

In contrast, the AP classifier, also introduced by

Bounhas et al. (2017), does not consider all possible

triplets from the training set when classifying a new

item 𝐷



⃗

. Instead, it restricts the search scope to a

smaller subset of candidate triplets. The AP classifier

first identifies examples most similar to the item to be

classified and narrows its search to pairs of examples

exhibiting the same degree of dissimilarity (measured

using Manhattan distance) as that between the new

item 𝐷



⃗

and one of its nearest neighbors. This

approach implicitly constructs triplets that are

analogically proportional to the new item across all

attributes.

Classification with the AP classifier involves an

additive aggregation of the truth values associated

with the pairs that can be analogically related to those

formed by the target item 𝐷



⃗

and its nearest neighbor.

Only pairs that yield a solvable analogical equation

for the classes are considered. The algorithm

proposed by Bounhas et al. (2017) achieves

comparable performance to earlier analogical

classifiers while demonstrating reduced average

computational complexity in both nominal and

numerical contexts. However, its evaluation has been

limited to UCI benchmark datasets for nominal and

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1210

numerical data. This study aims to extend its

application by evaluating the AP classifier's

effectiveness on real datasets for Arabic text

morphological disambiguation. Additionally, we

conduct a comparative analysis with several

competitive ML and DL algorithms.

The basic procedure of the AP classifier can be

summarized by the following steps:

- Search for triplets of examples (𝐴

⃗

,𝐵



⃗

,𝐶

⃗

) in the

training set s.t: 𝐶

⃗

is the nearest neighbor of 𝐷



⃗

- Solve the equation: Class(𝐴

⃗

):Class(𝐵



⃗

)::Class(𝐶

⃗

):X.

- If the previous analogical equation on classes has a

solution ℓ and if the analogical proportion 𝐴

⃗

:𝐵



⃗

∷

𝐶

⃗

:𝐷



⃗

is valid in a componentwise manner for each

attribute, then increment the score of ℓ as score(ℓ) =

score(ℓ)+ 1.

- Assign to 𝐷



⃗

the class label ℓ having the highest score

score(ℓ) as Class(𝐷



⃗

) = argmax

ℓ

(score).

Algorithm : AP classifier (Bounhas et al., 2017).

Input: k >1, S a training set, 𝐷



⃗

∉S a new instance to be classified

For each label

ℓ Do

Score(

ℓ) = 0

EndFor

For each 𝐶

⃗

in N

(𝐷



⃗

) Do

For each pair (𝐴

⃗

, 𝐵



⃗

) in |S|

If (Class (𝐴

⃗

) : Class(𝐵



⃗

) :: Class(𝐶

⃗

) : X has solution ℓ)

and (𝐴

⃗

:𝐵



⃗

∷𝐶

⃗

:𝐷



⃗

) then Score(ℓ) = Score(ℓ)+1

Endif

EndFor

Score* = max (Score(

ℓ))

if(Score* ≠ 0) then

if(unique(Score* , Score(

ℓ))) then

Class(𝐷



⃗

) = argmax

ℓ

(Score(ℓ))

else

Majority vote

Endif

else

unclassified

Endif

return Class(𝐷



⃗

)

The AP classifier's success depends on the presence of

class-solvable triplets in the training set that form an

analogical proportion with the new item to be

classified. If such triplets are absent, the classifier fails.

This limitation is more likely to occur with small

training sets. However, in our analysis of the Arabic

text morphological disambiguation datasets, this issue

did not arise. In cases where multiple candidate labels

are found i.e., the predicted label is not unique (see the

algorithm below), the ambiguity is resolved through a

majority vote among all possible candidate labels.

https://www.nongnu.org/aramorph/english/index.html

The disambiguation process for a given

ambiguous word is based on the context of the Arabic

text. For instance, we assume that the POS is the

morphological feature (MF) to be disambiguated. The

classification process relies on two preceding

attributes (POS-1 and POS-2) and two succeeding

attributes (POS+1 and POS+2) of this word.

5 EXPERIMENTAL RESULTS

In this section, we provide a brief overview of the test

collection in Section 5.1. The experimental scenario

is detailed in Section 5.2, followed by a comparative

study in Section 5.3, which highlights the efficiency

of the AP classifier in comparison to various ML and

DL classifiers.

5.1 Test Collection

We primarily aim to train the AP, ML and DL

classifiers by capturing morphological dependencies

using vocalized texts, and then we test these models

on non-vocalized texts. During training, we use the

morphological analyzer ARAMORPH

on vocalized

Arabic texts from Hadith to extract values for 14

morphological features. During the data pre-

processing step, a data transformation technique is

applied to convert imperfect data into perfect data

suitable for classical ML and DL classifiers (Elayeb

et al., 2022). The selected classifiers are subsequently

trained on vocalized Arabic texts and tested on non-

vocalized ones.

Table 1: Overview of the Arabic dataset.

Morphological Feature

(MF)

Size (Ko) Attributes Instances

POS 8.806 1961 1516

ADJECTIVE 470 105 1502

ASPECT 937 209 1501

CASE 931 209 1501

CONJUNCTION 472 105 1501

DETERMINER 1.186 266 1503

GENDER 696 157 1501

MODE 703 157 1501

NUMBER 926 209 1500

PARTICLE 938 209 1507

PERSON 932 209 1502

PREPOSITION 472 105 1502

VOICE 704 157 1501

PRONOUN 14.551 3329 1477

Table 1 presents an overview of the Arabic dataset

("Kunuz" corpus of Hadith texts (Bounhas et al.,

2010; 2011ab)) in terms of data size for the 14

Morphological Disambiguation of Texts Based on Analogical Proportions

1211

morphological features, including their total number

of attributes (having Boolean values) and instances.

5.2 Experimental Scenario

We apply a 10-fold cross-validation technique across

three application domains derived from six Hadith

books to assess the performance of the AP classifier,

three ML (SVM, Naïve Bayes, and Decision Tree)

and three DL classifiers (GRU, CNN and LSTM). For

each morphological feature, we compute the average

disambiguation rate over the (9+1) iterations.

To get the disambiguation rates, we follow these

steps: (i) we first analyze the vocalized texts and

record the correct morphological solutions; (ii)

second, we remove short vowels from the same texts;

(iii) third, we disambiguate the resulting texts using a

given classifier and we store the results; and (iv)

finally, we compare the two sets of results to compute

the disambiguation rate.

We experiment with the three ML classifiers

(SVM, NB and DT) using their optimized parameters

detailed in (Elayeb et al., 2022). Besides, the structure

of the Convolutional Neural Network (CNN) model

includes a dropout layer, followed by three CNN

layers with a kernel size of 5 and 128 filters. These

are followed by global max-pooling with default

parameters and another dropout layer. We also used

both LSTM and GRU algorithms. The LSTM

algorithm contains a single LSTM layer, while the

GRU algorithm consists of two GRU layers. This

configuration was determined through

experimentation to achieve optimal accuracy.

To ensure a fair comparison across all tested

classifiers, we also optimize the parameter k for the

AP classifier (k representing the number of nearest

neighbors 𝐶

⃗

considered for classifying an item).

Specifically, within each fold of the outer 10-fold

cross-validation, we first extract the training set.

Then, an inner 5-fold cross-validation is performed

on this training set to determine the optimal value of

k. The selected value of k from this initial step is

subsequently used to classify the test examples in the

corresponding outer fold. This procedure is repeated

for all folds in the outer cross-validation.

The classification results for the various

classifiers, presented in Table 2, correspond to the

optimal value of each tuned parameter.

5.3 Experimental Results and

Discussion

Table 2 summarizes the disambiguation rates of the 14

morphological features using AP, ML, and DL

classifiers. The results confirm that AP and ML

classifiers outperform DL classifiers in disambiguating

the following morphological features with similar

efficiency: ADJECTIVE (96.54%), ASPECT

(71.29%), CASE (56.16%), GENDER (57.16%),

MODE (99.40%), NUMBER (85.27%), and VOICE

(71.29%). Moreover, the AP classifier achieved an

almost identical rate (96.48%) if compared to SVM

and NB (96.75%) when disambiguating the

morphological feature PARTICLE. However, DL

classifiers achieved the best classification results for

CONJUNCTION (84.70%), DETERMINER

(68.40%), and PREPOSITION (84%). Furthermore,

CNN and LSTM emerged as the best classifiers for

disambiguating POS (78.00%) and PERSON

(61.70%), respectively. Conversely, the AP classifier

achieved the highest rate for disambiguating

PRONOUN (67.21%). Overall, the average

disambiguation rate of the AP classifier across the 14

morphological features is 74.80%, outperforming all

ML and DL classifiers. Notably, the data size of certain

morphological features poses challenges for some

classifiers (e.g., POS for ML classifiers and

PRONOUN for DL classifiers). For example, POS

includes 1,961 attributes, while PRONOUN includes

3,329 attributes. ML classifiers, in particular, struggle

to process this data, even with WEKA's maximum

memory allocation of 2,020 MB. To address this

limitation, randomly selected subsets of the data are

used instead of the entire dataset. However, this

method leads to a reduction in disambiguation

accuracy for these extensive morphological features if

compared to smaller ones.

Furthermore, we observe that certain classifiers

produced similar or identical outcomes for specific

morphological features. This can be explained by the

limited number of possible values for these features

(fewer than six) (see for example the results for the

Adjective feature with only two possible classes). In

contrast, other features yield diverse results across

different classifiers. For example, the feature

PRONOUN includes 64 possible class values. These

observations highlight that effectively optimizing

classifier parameters plays a crucial role in improving

their performance when disambiguating Classical

Arabic texts. Moreover, managing small corpora and

text collections with a large number of attributes

remains one of the primary challenges of existing DL

classifiers. For instance, the GRU algorithm

demonstrates a low disambiguation rate (20.20%) for

the morphological feature PRONOUN, which

comprises 3,329 attributes. These findings align with

the conclusions of Elnagar et al. (2020), who also

tested DL algorithms for Arabic text classification.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1212

Table 2: Disambiguation rates of 14 morphological features using AP, ML and DL classifiers.

Morphological Feature

(MF)

ML classifiers DL classifiers

Analogical

classifier

SVM NB DT GRU CNN LSTM AP

POS 29.67% 36.21% 48.94% 73.00% 78.00% 31.20% 56.01%

ADJECTIVE 96.54% 96.54% 96.54% 96.00% 96.10% 96.00% 96.54%

ASPECT 71.29% 71.29% 71.29% 69.70% 70.10% 73.00% 71.29%

CASE 56.16% 56.16% 56.16% 21.50% 51.50% 58.80% 56.16%

CONJUNCTION 83.08% 83.08% 83.08% 84.70% 84.70% 84.70% 83.08%

DETERMINER 64.20% 64.20% 64.20% 68.40% 68.40% 68.40% 64.13%

GENDER 57.16% 57.16% 57.16% 55.10% 55.10% 55.10% 57.16%

MODE 99.40% 99.40% 99.40% 99.30% 99.30% 99.30% 99.40%

NUMBER 85.27% 85.27% 85.27% 45.20% 45.20% 85.60% 85.27%

PARTICLE 96.75% 96.75% 96.68% 43.00% 93.00% 94.30% 96.48%

PERSON 60.25% 60.25% 60.25% 60.80% 60.80% 61.70% 60.25%

PREPOSITION 82.89% 82.89% 82.89% 84.00% 84.00% 84.00% 82.89%

VOICE 71.29% 71.29% 71.29% 13.00% 70.90% 73.00% 71.29%

PRONOUN 63.04% 60.19% 62.56% 20.20% 58.10% 58.10% 67.21%

Avera

e 72.64% 72.90% 73.97% 59.60% 72.10% 73.10% 74.80%

The AP classifier appears to be less sensitive to

dataset size, even when managing a large number of

attributes or classes. Notably, it achieves the highest

disambiguation rate for the PRONOUN dataset

compared to all other classifiers. This highlights, once

again, the AP classifier's efficiency in handling multi-

class classification tasks involving numerous

attributes, even under conditions of data scarcity

(Bounhas et Prade, 2023). This efficiency is achieved

through the use of triplets of examples from the

training set, which form an analogical proportion with

the item being classified. Consequently, the triplet-

based approach serves as a method for augmenting

sparse data, allowing the classifier to draw reliable

conclusions about the item by aggregating scores

across various triplets.

6 CONCLUION

Arabic morphological disambiguation remains one of

the major challenges in several domains, including

machine translation, information retrieval, speech

recognition and synthesis, educational tools, and text-

to-speech systems. Solving the problem of ambiguity

in Arabic texts with high accuracy can be highly

beneficial for these application areas. For this purpose

and given the success of analogical proportions in

summarizing Arabic texts and in classifying both

structured nominal or numerical data and unstructured

data (such as Arabic text classification), we aim to

investigate the efficiency of an AP classifier in

disambiguating Classical Arabic texts. We compare

the performance of the analogy-based algorithm to a

set of well-established ML and DL classifiers. The

results demonstrate the competitive performance of the

AP classifier in terms of the average disambiguation

rate across the 14 morphological features.

Furthermore, the AP classifier exhibits reduced

complexity compared to analogical classifiers that take

into account all triplets, while maintaining accuracy

that is either better than or, in many cases, equivalent

to some ML and DL classifiers (Bounhas et al, 2024.a).

Despite its efficiency, the AP classifier requires

further investigation and enhancement. First, it is

important to test it using modern Arabic text

collections, such as TreeBank. Second, we aim to

expand the AP classifier to accommodate regional

dialects with unique morphological structures.

Finally, we believe that the accuracy of Arabic

morphological disambiguation can be improved by

incorporating advanced linguistic context, including

integration with syntax and semantics. Additionally,

we propose leveraging unsupervised learning

techniques to reduce reliance on annotated datasets,

focusing on the development of unsupervised or

semi-supervised methods.

REFERENCES

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

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

and Disambiguation Using a Possibilistic Classifier. In

Proc. 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 Proc. 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

Proc. of AICCSA-2018, pp. 1-6, Aqaba, Jordan, IEEE

Computer Society.

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

Possibilistic Morphological Disambiguation of

Morphological Disambiguation of Texts Based on Analogical Proportions

1213

Structured Hadiths Arabic Texts Using Semantic

Knowledge. In Proc. 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 Proc. 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 Proc. 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 Knowl. Eng.,

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. Comput.

Speech Lang., 33(1):67-87.

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

Toward a computer study of the reliability of Arabic

stories. J. Assoc. Inf. Sci. Technol., 61(8):1686–1705.

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

ArabOnto: Experimenting a New Distributional

Approach for Building Arabic Ontological Resources.

Int. J. Metadata Semant. Ontologies, 6(2):81-95.

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

Organizing Contextual Knowledge for Arabic Text

Disambiguation and Terminology Extraction. Knowl.

Org., 38(6):473–490.

Bounhas, M., Elayeb, B. (2019). Analogy-based Matching

Model for Domain-specific Information Retrieval. In

Proc. of ICAART-2019, Vol. 2, pp. 496-505, Prague,

Czech Republic, SciTePress.

Bounhas, M., Elayeb, B., Chouigui, A., Hussain, A.,

Cambria, E. (2024). Arabic text classification based on

analogical proportions. Expert Syst. J. Knowl. Eng.

41(10).

Bounhas, M., Prade, H. (2023). Analogy-based classifiers:

An improved algorithm exploiting competent data

pairs. Int. J. Approx. Reason., 158: 108923.

Bounhas, M., Prade, H. (2024). Revisiting analogical

proportions and analogical inference. Int. J. Approx.

Reason. 171: 109202.

Bounhas, M., Prade, H., Richard, G. (2017). Analogy-based

classifiers for nominal or numerical data. Int. J. Approx.

Reason., 91: 36-55.

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

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

Morpho-Lexical Disambiguation of Arabic. J. Inf.

Process. Syst., 12(3):358–380.

Daoud, D. (2009). Synchronized Morphological and

Syntactic Disambiguation for Arabic. Adv. Comput.

Linguistics, 41:73–86.

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

Using Dependency Grammar. In Proc. of TALN-2009,

Senlis, France, ATALA.

Diab, M. T. (2007). Improved Arabic Base Phrase Chunking

with a New Enriched POS Tag Set. In Proc. of Semitic-

2007, pp. 89–96, Stroudsburg, PA, USA. ACL.

Dubois, D., Prade, H., Richard, G. (2016). Multiple-valued

extensions of analogical proportions, Fuzzy Sets Syst.,

292:193–202.

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

Review. Artif. Intell. Rev., 52(4):2475-2532.

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

Review. In Proc. of AICCSA-2021, pp. 1-8, Tangier,

Morocco, IEEE Computer Society.

Elayeb, B., Ayed, R. (2022). Socio-Semantic Information

Retrieval of Structured Arabic Texts. In Proc. of

AICCSA-2022, pp. 1-8, Abu Dhabi, UAE, IEEE

Computer Society.

Elayeb, B., Ayed, R. (2023). Analogical Text Mining:

Application to Arabic Text Summarization and

Classification. In Proc. of AICCSA-2023, pp. 1-8, Giza,

Egypt, IEEE Computer Society.

Elayeb, B., Ben Khiroun, O. (2023). SPEEDSER: A

Possibilistic System for Query Disambiguation,

Expansion and Translation. Int. J. Inf. Technol. Decis.

Mak. DOI: 10.1142/S0219622023500499

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

Information Retrieval: A Review. ACM Trans. Asian

Low Resour. Lang. Inf. Process., 15(3):18:1-18:44.

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

(2020). Automatic Arabic Text Summarization Using

Analogical Proportions. Cogn. Comput., 12(5):1043-1069.

Elayeb, B., Ettih, M.F., Ayed, R. (2022). Experimenting

machine-learning algorithms for morphological

disambiguation of Arabic texts. In Proc. of ICAART-

2022, Vol. 3, pp. 851-862, SciTePress.

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

(2009). Towards an intelligent possibilistic web

information retrieval using multiagent system. Interact.

Technol. Smart Educ., 6(1):40–59.

Elnagar, A., Debsi, R. A., Einea, O. (2020). Arabic text

classification using deep learning models. Inf. Process.

Manag., 57(1), 102121.

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

Through Full Morphological Tagging. In Proc. of HLT-

NAACL 2007: Short Papers, HLT-NAACL-Short’07,

pp. 53–56, Stroudsburg, PA, USA. ACL.

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

Proc. of the NAACL-2001, Pennsylvania, USA.

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

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

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

Tool for Morphological Analysis and Disambiguation

of Arabic. In Proc. of LREC-2014, pp. 1094-1101,

Reykjavik, Iceland, ELRA.

Prade, H., Richard, G. (2010). Reasoning with logical

proportions. In Proc. of KR-2010, pp.545–555,

Toronto, AAAI Press.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1214