LLMs Based Approach for Quranic Question Answering

Zakia Saadaoui

1,2 a

, Ghassen Tlig

3 b

and Fethi Jarray

1,2 c

LIMTIC Laboratory, UTM University, Tunisia

Higher Institute of Computer Science of Medenine, Tunisia

Ecole Sup

erieure d’Electronique de l’Ouest, Paris, France

Keywords:

Quranic QAS, LLM, Prompt Engineering.

Abstract:

This paper addresses a prominent research gap in Quranic question answering, where current methodologies

face challenges in capturing the nuanced aspects of inquiries related to the Quran. By presenting an innovative

approach that utilizes Large Language Models (LLMs), including GPt4, Bart, and LLAMA, we aim to over-

come these limitations, improving the clarity and precision of responses to Quranic queries. The evaluation of

the proposed Quranic Question Answering System, using the F1 score metric, demonstrates encouraging re-

sults in comprehending and addressing various Quranic queries. Notably, the application of a chain of thought

with Bart achieves an impressive F1-score of 95%. This research offers a distinctive perspective on Quranic

question answering through the integration of LLMs.

1 INTRODUCTION: QA

SYSTEMS

Question Answering (QA) systems play a pivotal role

in information retrieval and can be broadly catego-

rized into two main types: Open Domain and Closed

Domain. In Open Domain QA, models operate with-

out constraints to a speciﬁc domain, drawing infor-

mation from diverse sources like books, the internet,

Wikipedia, tables, graphs, and knowledge bases. On

the other hand, Closed Domain QA systems are tai-

lored for speciﬁc domains, such as legal or health-

care documents. Both Open Domain and Closed Do-

main QA systems can be further classiﬁed into open

and closed book QA systems. Addressing the com-

plexities of the Quran presents unique challenges for

question answering systems. Existing methods of-

ten struggle to grasp the intricacies of Quranic lan-

guage and context, leading to incomplete or inac-

curate responses. This paper addresses this critical

gap in Quranic AI, proposing innovative approaches

for more comprehensive and insightful answers. The

complexity of Quranic questions requires advanced

natural language processing (NLP) techniques such

as the utilization of Large Language Models (LLMs).

https://orcid.org/0000-0001-8695-2034

https://orcid.org/0000-0002-1911-9170

https://orcid.org/0000-0002-5110-1173

Large Language Models, such as GPt4, Bart, and

LLAMA, have demonstrated exceptional capabili-

ties in understanding and generating human-like text

across various domains. However, their application to

answering Quranic questions remains relatively unex-

plored.

In this paper, we present a novel approach that

leverages LLMs to enhance the comprehension and

accuracy of Quranic question answering. Speciﬁcally,

we focus on GPt4 tailoring their capabilities to the

unique challenges posed by Quranic text. By doing

so, we aim to ﬁll the existing research gap and con-

tribute to the advancement of Quranic question an-

swering systems. The proposed Quranic Question

Answering System is rigorously evaluated using the

F1-score metric, providing insights into its effective-

ness in comprehending and addressing a diverse array

of Quranic queries. Notably, employing a chain of

thought with Bart achieves an exceptional F1-score

of 95%, underscoring the potential of this approach.

1.1 Architecture of Arabic QAS

A Question Answering System (QAS) typically con-

sists of three several components :question analy-

sis, passage retrieval, and answer extraction. Ques-

tion Analysis module analyzes the structure of the

question, identiﬁes the question type : factual, (e.g.,

where, when, who, what) or a taxonomy manually

112

Saadaoui, Z., Tlig, G. and Jarray, F.

LLMs Based Approach for Quranic Question Answering.

DOI: 10.5220/0013012900003825

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

In Proceedings of the 20th International Conference on Web Information Systems and Technologies (WEBIST 2024), pages 112-118

ISBN: 978-989-758-718-4; ISSN: 2184-3252

deﬁned by linguistic experts. Based on the question

analysis, this module generates a structured query or

representation that can be used to retrieve relevant in-

formation from a knowledge source. A retrieval is a

component that retrieves relevant passages or docu-

ments from a knowledge base or a corpus. An An-

swer Extraction this is module identiﬁes potential an-

swer candidates within the retrieved passages or docu-

ments. Answer Ranking and Selection: Once answer

candidates are identiﬁed, they may be ranked based

on their relevance and conﬁdence scores. The top-

ranked answers are selected and returned to the user.

In some cases, A QAS needs a natural language Un-

derstanding and Generation module to reﬁne the an-

swer through additional processing such as summa-

rization or paraphrasing to ensure clarity and coher-

ence and to generate a ﬁnal answer to be formatted

and presented to the user in natural language, along

with any relevant contextual information or sources.

the progress in NLP techniques, particularly with

the advent of deep learning and transformer-based

models like BERT, GPT, and their variants, has sig-

niﬁcantly improved the accuracy and capabilities of

QASs. These models have led to innovations in var-

ious components of the QAS architecture, it enables

end-to-end formability. For question analysis, work

is developing neural classiﬁers to determine question

types. For question analysis, work is underway to de-

velop neural classiﬁers to determine question types.

For example, the CNN-based model and the LSTM-

based model are used respectively to classify given

questions, and both achieve competitive results. For

document retrieval, methods based on dense repre-

sentation have been proposed to solve the ”term mis-

match” problem, which has long hampered retrieval

performance. Unlike traditional methods such as TF-

IDF and BM25 , deep search methods learn to en-

code questions and documents in a latent vector space

where text semantics beyond term matching can be

measured.each document and question independently

in dense vectors, and the similarity score can be cal-

culated.the same for answer extraction neural mod-

els can also be applied.So the new architecture take

a query in natural language to be answered by a QA

model composed with two main components the re-

triever, this component is responsible for extracting

relevant documents or passages from a large corpus

of texts according to the input query. It uses tech-

niques like TF-IDF Term Frequency-Inverse Docu-

ment Frequency, Best Matching 25, or neural-based

approaches to efﬁciently retrieve a subset of docu-

ments likely to contain the answer to the user’s query.

Once the relevant passages have been retrieved by the

extractor, the reader comes into play. The reader is

tasked with understanding the retrieved text and ex-

tracting precise answers to the user’s question from

these passages. This often involves the use of natural

language processing (NLP) techniques, such as au-

tomatic comprehension models like BERT (Bidirec-

tional Encoder Representations from Transformers)

or its variants, which are trained speciﬁcally for tasks

such as answering questions by extraction.

Figure 1: Architecture of QAS based on Retriever-Reader.

We propose in table 1 a new vision for building

of QAS deponds in the model of QA that composed

from the retriever and the reader . that can give a new

classiﬁcation of QAS shown in 2: a open extractive

book based on information retriever and and a ma-

chine reading comprehension to read answers from a

given context and a open abstractive book that the an-

swer is a text generated from a QA model.

Figure 2: Approaches to Question-Answering.

2 RELATED WORK

Arabic Question Answering System (QAS) presents a

number of challenges and peculiarities compared to a

traditional QAS using other languages such as the lin-

guistic complexity: The Arabic language has a com-

plex grammatical and syntactic structure, with partic-

ular declension and conjugation rules. This makes au-

tomatic comprehension of questions and ﬁnding an-

swers more difﬁcult than in languages with simpler

structures and the scarcity of linguistic resources: Un-

like some Western languages, for which there are

vast annotated datasets, lexicons and linguistic re-

sources.in the last the sensitivity to cultural contexts:

Questions asked in Arabic may be inﬂuenced by spe-

ciﬁc cultural and religious references. QAS must be

sensitive to these contexts in order to provide appro-

priate answers.

LLMs Based Approach for Quranic Question Answering

113

Table 1: Building QAsystem.

ExtractiveQA Open Generative Closed Generative

the use of Context yes yes no

way of the answer is obtained Extraction Generation Generation

2.1 Arabic Question Answering System

Question Answering (QA) tasks can be grouped into

different categories according to the type of ques-

tion asked: factoid, list, deﬁnition and non-factoid.

Factual questions generally concern speciﬁc enti-

ties. Questions beginning with words such as When,

Where, How much, What and Who are other exam-

ples of this type. In Arabic, these types of questions

are often used in research.Various Arabic QAS have

been proposed in the literature. QARAB is a system

proposed by (Hammo et al., 2002) which takes fac-

toid Arabic questions and attempts to provide short

answers. ArabiQA proposed by Benajiba et al in (Be-

najiba et al., 2007) its research work is interested in

modern Arabic language. It answers factoid questions

based on named entity recognition.QASAL (Brini

et al., 2009)is designed to answer factoid and def-

inition questions.AQUASYS (Bekhti and Al-Harbi,

2013a)is designed to answer Arabic questions related

to named entities of types person, location, organiza-

tion, time, quantity, The authors assumed that the an-

swer is a short passage.IDRAAQ (Abouenour et al.,

2012)is an OpenQAS designed to enhance the the

quality of retrieved passages.JAWEB (Kurdi et al.,

2014)is based on AQUASYS (Bekhti and Al-Harbi,

2013b). JAWEB provides a web interface to the sys-

tem, which is an additional support for Arabic lan-

guage presentation in web browsers. The system

scored 100% recall and 80% precision. LEMAZA

(Azmi and Alshenaiﬁ, 2017)is built to answer Ara-

bic why-questions.this system achieved about 72.7%

Recall, 79.2% Precision. The SOQAL system, as de-

scribed by (Mozannar et al., 2019a), is notable for be-

ing the ﬁrst Arabic Open Question Answering Sys-

tem (OpenQAS) to employ a neural approach. It

comprises two primary components: a retriever and

a reader. for the retriever it uses TF-IDF technique to

retrieve ﬁrst a set of documents related to the ques-

tion .for The reader it is a neural reading comprehen-

sion model based on BERT.SOQAL achieved 20,7%,

42,5%, and 51,7% in exact match score, f1 score, and

sentence match score, respectively. Arabic QA4MRE

presnted by Trigui et al in (Trigui et al., 2012) in-

troduced the Arabic language for the ﬁrst time to

CLEF. The system proposes a new approach that al-

lows questions with multiple answer choices to be an-

swered from short Arabic texts.

2.2 Arabic QA Datasets

In this context we present a set of arabic Question

answering datasets we propose to classify them into

categories as shown in ﬁgure 2 .

ARCD: a open book dataset Crowd-sourced com-

posed by 1,395 questions based on 465 paragraphs

from 155 articles presented by (Mozannar et al.,

2019b)

TYDIQA: is a closed book question answering

dataset covering 11 typologically diverse languages

with 204K question-answer pairs presented in (Clark

et al., 2020).

Arabic-SQuAD: a open book dataset for extractive

QA system composed by 48,344 questions on 10,364

paragraphs that are translated from English SQuAD

presented in (Mozannar et al., 2019b)

MLQA: is MultiLingual Question Answering)

,a benchmark dataset for evaluating cross-lingual

question answering performance.It is a extractive QA

proposed by(Clark et al., 2020)

SOQAL: is a open book dataset for extractive QA

based on : a document retriever using a hierarchical

TF-IDF approach and a neural reading compre-

hension model using the pre-trained bi-directional

transformer BERT (Clark et al., 2020)

So2al-wa-Gwab: A New Arabic Question-

Answering Dataset Trained on Answer Extraction

Models presented in (Al-Omari and Duwairi, 2023)

CQA-MD: a medical Arabic corpus for closed book

was proposed by (Nakov et al., 2019). The corpus

contains over 100k questions-answers pairs collected

from Arabic Medical websites.

2.3 Quranic Question Answering

System

Many studies have building systems for locating an-

swers to Quranic inquiries from the Holy Quran

. These investigations generally employed differ-

ent techniques. Abdelnasser et al.(2014) in (Abdel-

nasser et al., 2014)have focused in retrieval tech-

niques to developing a open extractive book system

named Al-Bayan.This system taked the question as

input and then retrieves a Quran verse that includes

the answer based on ontology by computing the co-

sine semantic similarity between the question and the

concept vectors.Other researchers try to extract an-

swers from Hadith using the same techniques, such

WEBIST 2024 - 20th International Conference on Web Information Systems and Technologies

114

Table 2: Quranic QA datasets.

Dataset Description

A corpus of Quran

and tafsir (Abdel-

nasser et al., 2014)

A Collection of dataset from a corpus of the quran and

tafsir(Ontology)

Surah al-

baquarah(Ahmad

et al., 2016)

Most popular English translation of the Quran of Abdullah Yusuf Ali.

Quran DB (Alqah-

tani and Atwell,

2018)

Arabic-English Quran ontologies from different datasets related to

Al Quran.

Ayatec (Malhas and

Elsayed, 2020)

207 questions (with their corresponding 1,762 answers) covering 11

topic categories of the Holy Qur’an

QRCD (Malhas and

Elsayed, 2022)

Extractive open book QA for Qur’anic Reading Comprehension

Dataset

QUQA (Alnefaie

et al., 2023)

2,189 questions, classiﬁed as 1778 single-answer and 411 multiple-

answer questions

as the work of (Maraoui et al., 2021) with an accu-

rancy of 92In (Hamoud and Atwell, 2016) recom-

mend building a open extractive system on a knowl-

edge base of related domains to answer all kinds

of questions about the Qur’an. They constructed a

corpus of 1,500 questions and their answers. The

dataset included different types of questions.The sys-

tem is based on matching the user query to the ques-

tions in the dataset in order to ﬁnd the most rel-

evant question and display its answer. This sys-

tem demonstrated a precision of 79 % and a recall

of 76%. The performance of this approach is af-

fected by the size and variety of the corpus. In

recent years, several studies have focused on the

use of pre-trained models for the Qur’anic Machine

Reading Comprehension (MRC) task. The authors

used the Qur’anic Reading Comprehension Dataset

(QRCD).(ElKomy and Sarhan, 2022) to build a ex-

tractive QA for open book. they tested ﬁve pre-trained

models: Arabic BERT (ARBERT), AraBERTv02-

Base, AraBERTv02-Large, Masked Arabic BERT

(MAR- BERT) and QCRI [Qatar Computing Re-

search In- stitute] Arabic and Dialectical BERT

(QARiB)-Base. Their system yielded the following

results: 0.27 Exact Match (EM), 0.50 F1@1 and 0.57

partial Reciprocal Rank (pRR). (Ahmed et al., 2022)

augmented the QRCD with 657 questions and trained

the AraElectra model. The system yielded the fol-

lowing results: 0.24 EM, 0.51 F1@1 and 0.55 pRR.

Malhas and Elsayed (2022) in (Malhas and Elsayed,

2022) interested in reading comprehension to build

an open extractive book , they pre-trained the Arabert

model on data in classical Arabic and then they ﬁne

tunned it on data from the QRCD dataset. This model,

called CL Arabert, outperformed Arabert with a par-

tial precision of 0.51. With the advent of large lan-

guage models, the focus turned to the use of these

models in NLP tasks, particularly In Question An-

swering system.In (Alnefaie et al., 2023) introduced

the use of the GPT family to generate answers to ques-

tions from the QUQA dataset, they made just the ex-

periments and they use a manuel evaluation Perfor-

mance evaluations of this system resulted in F1=0.26

for the text portion of the GPT4 generated answers

and F1=0.32 for the evaluation of the Quranic series

portion of the GPT-4 answers.

2.4 Quaranic QA Datasets

We present a summary of Qur’anic datasets in the ta-

ble 2.

2.5 Large Language Models

Large Language Models (LLMs) such as the GPT

series mentioned by research (Radford et al., 2018)

.GPT’s progression has already gone through sev-

eral genera- tions: GPT-1 (Radford et al., 2018),

GPT-2(Radford et al., 2019) , GPT-3 (Brown et al.,

2020),and GPT-4 (OpenAI, 2023).they have been at

the forefront of recent NLP research following their

ability to generate consistent and contextually ac-

curate textual results. These models have demon-

strated impressive performance in natural language

processing (NLP) tasks ranging from summarization

and translation to question answering.

the strategy we are going to implement in this article

to question LLM is the prompt engineering in order to

guide our Quranic question answering system to pro-

duce good results.

LLMs Based Approach for Quranic Question Answering

115

Figure 3: LLM for QA.

2.6 Prompt Engineering

Prompt engineering is the practice of designing ef-

ﬁcient inputs, or prompts, to achieve desired results

from AI language models such as LLMs.It has many

objectives such as:

• Improved performance: Well-designed prompts

can help AI models produce more accurate, rel-

evant and consistent results. This is particularly

important in applications where precise, nuanced

or contextual responses are required.

• User experience: Effective prompt engineering

contributes to a better user experience by enabling

AI models to understand the user’s intent more

precisely, and to respond in a more useful and

compliant way to their expectations.

• Risk mitigation: Carefully designed prompts can

reduce the likelihood of AI models generating

inappropriate content by guiding them towards

safer, more context-aware results.

• Domain adaptation: Prompt engineering enables

AI linguistic models to be adapted more effec-

tively to speciﬁc domains or applications, even

when they have been trained beforehand on gen-

eral linguistic data ,in our case, this is Quranic

question-answer systems.

2.6.1 Techniques of prompt Engineering

there is a speciﬁc types of prompting depends in con-

text of application of this technique.

• Zero shot Prompting: With this technique we

do not provide the LLM with examples of texts in

parallel to get better answers.

• Few−shot Prompting: is a technique where the

input consists of a natural language query or in-

struction, followed by one or more examples of

the desired output.

• Chain of Thought: Wei et al (2022) proposed the

”chain of thought” (CoT) prompt, which achieves

complex reasoning abilities through intermediate

reasoning steps.

• Self Consistency: Introduced by Wang et al

(2022), self-consistency aims to ”replace naive

decoding in chain-of-thought prompts”.It im-

proves the performance of chain-of-thought

prompting in tasks involving arithmetic reason-

ing.

3 PROPOSED APPROACH

Our proposed approch consists in building a Open-

book abstractive Quranic question answering sys-

tem. we will build a reader-retriever model and ex-

plore its capabilities in providing accurate and well-

sourced answers to our Quranic questions based on

the technique of zero shot prompting using GPT-4-

turbo LLM tested the QUQA dataset. Our approach is

to ameliorate the output of the retriever and restrict the

relevant passages to a verses of the Qur’an and gener-

ate answers in a speciﬁc format where it will be easy

to apply automatic evaluation. The QUQA contains

three kinds of question. The QUQA contains three

types of questions: a 103 conﬁrmation questions that

the answer can be yes or no those question begin with

Éë , a 1,621 descriptive questions started with

Q»





@ , @



XAÓ , @



XAÖ



J



» , Õ» , AÓ

their answers can be reason explanation or a deﬁnition

and 465 factoid questions begin with



æÓ ,



áK





@ ,



áÓ

with a answer type a location, a duration or a name

of person . The version used for this study was GPT-

4-turbo, it is latest version of the GPT at the time of

our research. This model is used directly and does

not require any ﬁne-tuning process but it needs guid-

ance We used the Python API to retrieve answers from

the model. Google Colab was used to run our experi-

ments.

3.1 Metrics of Evaluation

The output generated by GPT-4-turbo ,our LLM in

response to our provided questions produces natural-

sounding text, often including several series of

Quranic verses, These series may vary in length, com-

prising one or more verses. Consequently, we con-

ducted separate evaluations for the textual answers

and the Quranic verses themselves.The evaluation

of the Quranic verses involved both automated pro-

cesses.by prompting the system to display a ranked

list of the Quranic verses in the answer, we extract

a list of predictions and compare them with differ-

ent references answers mentioned as gold responses.

In order to evaluate the generated responses, we used

the evaluation metrics mentioned in the research work

WEBIST 2024 - 20th International Conference on Web Information Systems and Technologies

116

of (Malhas and Elsayed, 2022), we used F1 which

is a measure of a test’s accuracy that considers both

the precision and recall of the test to compute the

score ,To compute it, we measured the F1 between

each series of verses and the golden answer or the list

of golden answer for the question multi answers and

then took the average. The Exact matching EM and

the F1@1 are calculated in relation to the answer at

rank 1.for F1@1 we take the maximum if the ques-

tion has many golden answers.

4 RESULTS

In this section we present the results of our exper-

iments, we try several prompts in the query of our

model, the different prompts generate different re-

sponses. We carried out three large-scale experiments

on the QUQA dataset, dividing the experiments by

question type. Each large-scale experiment was

repeated several times, each time modifying the

prompt system.In the ﬁrst experimentation we set

just the user prompt regardless of the question asked,

then in the remaining experimentation we deﬁned a

system prompt to improve the evaluation results.The

results of the ﬁrst experiment are shown in the table

3, while the results of the second experiment, in

which the prompt system was mentioned .

Example 1 of prompt of system :



áÓ



éJ



ÓCB@ ©k

@QÖ

@ I

k È@



ñË@



á« I

Õç'



QºË@







®Ë@

are shown in table 4.

The results of evaluation of Example 2 of prompt of

system:



áÓ



éËX





éÖ





¯ Q»



hðQ¢Ö

@ È@



ñË@

úÎ

« I

Õç'



QºË@







®Ë@

are shown in table 5

Table 3: Results of ﬁrst experimentation prompt of user =

Question.

Fscore EM F@1

Conﬁrmation 0,17 0.11 0.19

Factoid 0.25 0.15 0,23

Descriptive 0,31 0,19 0,25

All 0,24 0,15 0,22

Table 4: Results of prompt 1 of System.

Fscore EM F@1

Conﬁrmation 0,22 0.13 0.21

Factoid 0.25 0.23 0,32

Descriptive 0,33 0,20 0,28

All 0,26 0,18 0,27

Table 5: Results of evaluation of our prompt2 of System.

Fscore EM F@1

Conﬁrmation 0,34 0.20 0.30

Factoid 0.38 0.25 0,32

Descriptive 0,40 0,28 0,35

Our evaluation of 0,37 0,24 0,32

(Alnefaie et al., 2023) 0,32 0,19 0,26

5 ANALYSE AND DISCUSSION

The QUQA dataset contains three kinds of question:

a 103 conﬁrmation questions that the answer can be

yes or no those questions begin with Éë , a 1,621 de-

scriptive questions started with

Q»





@ , @



XAÓ , @



XAÖ



J



» , Õ» , AÓ

their answers can be reason explanation or a deﬁnition

and 465 factoid questions begin with



æÓ ,



áK





@ ,



áÓ

with a answer type a location, a duration or a name of

person.

This section analyses and discusses the results of the

evaluation of our system based on the prompting tech-

nique. In ﬁrst case ,the prompt contains just the ques-

tion without any context In the Second case, the re-

sults obtained for the conﬁrmation questions are im-

proved by the quality of the prompting. In this case,

it is important to guide the model to ﬁnd the Qur’anic

verses corresponding to the answer. In the third case

,Our prompt system 2 is a conditional text genera-

tion system based on prompts.This technique enables

the generation of controlled and contextually relevant

text, making it useful for the evaluation system.

6 CONCLUSION

In this paper, we have evaluated a Quran question an-

swering system based on LLM specially the GPT4-

turbo version ,our main objective is to know how the

evaluation results as evolve as a function of prompt-

ing. As a future work, we aim a generalization to Con-

tinuous our experimentation using other techniques of

prompting like few shot or chain of taught combined

LLMs Based Approach for Quranic Question Answering

117

with other techniques like Retrieval Augmented Gen-

eration (RAG) for LLMs on other datasets.

REFERENCES

Abdelnasser, H., Ragab, M., Mohamed, R., Mohamed, A.,

Farouk, B., El-Makky, N. M., and Torki, M. (2014).

Al-bayan: an arabic question answering system for the

holy quran. pages 57–64.

Abouenour, L., Bouzoubaa, K., and Rosso, P. (2012).

Idraaq: New arabic question answering system based

on query expansion and passage retrieval.

Ahmad, N. D., Bennett, B., and Atwell, E. (2016).

Semantic-based ontology for malay qur’an reader.

Ahmed, B., Saad, M., and Refaee, E. A. (2022). Qqateam

at qur’an qa 2022: Fine-tunning arabic qa models for

qur’an qa task. pages 130–135.

Al-Omari, H. and Duwairi, R. (2023). So2al-wa-gwab: A

new arabic question-answering dataset trained on an-

swer extraction models. ACM Transactions on Asian

and Low-Resource Language Information Processing,

22(8):1–21.

Alnefaie, S., Atwell, E., and Alsalka, M. A. (2023). Is gpt-4

a good islamic expert for answering quran questions?

pages 124–133.

Alqahtani, M. and Atwell, E. (2018). Annotated corpus of

arabic al-quran question and answer.

Azmi, A. M. and Alshenaiﬁ, N. A. (2017). Lemaza: An

arabic why-question answering system. Natural Lan-

guage Engineering, 23(6):877–903.

Bekhti, S. and Al-Harbi, M. (2013a). Aquasys: A question-

answering system for arabic. 25(6):19–27.

Bekhti, S. and Al-Harbi, M. (2013b). Aquasys: A question-

answering system for arabic. 25(6):19–27.

Benajiba, Y., Rosso, P., and Lyhyaoui, A. (2007). Imple-

mentation of the arabiqa question answering system’s

components. pages 3–5.

Brini, W., Ellouze, M., Mesfar, S., and Belguith, L. H.

(2009). An arabic question-answering system for fac-

toid questions. pages 1–7.

Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D.,

Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G.,

Askell, A., et al. (2020). Language models are few-

shot learners. Advances in neural information pro-

cessing systems, 33:1877–1901.

Clark, J. H., Choi, E., Collins, M., Garrette, D.,

Kwiatkowski, T., Nikolaev, V., and Palomaki, J.

(2020). Tydi qa: A benchmark for information-

seeking question answering in ty pologically di verse

languages. Transactions of the Association for Com-

putational Linguistics, 8:454–470.

ElKomy, M. and Sarhan, A. M. (2022). Tce at qur’an qa

2022: Arabic language question answering over holy

qur’an using a post-processed ensemble of bert-based

models. arXiv preprint arXiv:2206.01550.

Hammo, B., Abu-Salem, H., Lytinen, S. L., and Evens, M.

(2002). Qarab: A: Question answering system to sup-

port the arabic language.

Hamoud, B. and Atwell, E. (2016). Quran question and

answer corpus for data mining with weka. pages 211–

216.

Kurdi, H., Alkhaider, S., and Alfaiﬁ, N. (2014). Develop-

ment and evaluation of a web based question answer-

ing system for arabic language. Computer science &

information technology (CS & IT), 4(02):187–202.

Malhas, R. and Elsayed, T. (2020). Ayatec: building a

reusable verse-based test collection for arabic ques-

tion answering on the holy qur’an. ACM Transac-

tions on Asian and Low-Resource Language Informa-

tion Processing (TALLIP), 19(6):1–21.

Malhas, R. and Elsayed, T. (2022). Arabic machine

reading comprehension on the holy qur’an using cl-

arabert. Information Processing & Management,

59(6):103068.

Maraoui, H., Haddar, K., and Romary, L. (2021). Ara-

bic factoid question-answering system for islamic sci-

ences using normalized corpora. Procedia Computer

Science, 192:69–79.

Mozannar, H., Hajal, K. E., Maamary, E., and Hajj, H.

(2019a). Neural arabic question answering. arXiv

preprint arXiv:1906.05394.

Mozannar, H., Hajal, K. E., Maamary, E., and Hajj, H.

(2019b). Neural arabic question answering. arXiv

preprint arXiv:1906.05394.

Nakov, P., M

arquez, L., Moschitti, A., and Mubarak, H.

(2019). Arabic community question answering. Nat-

ural Language Engineering, 25(1):5–41.

OpenAI, R. (2023). Gpt-4 technical report. arxiv

2303.08774. View in Article, 2(5).

Radford, A., Narasimhan, K., Salimans, T., Sutskever, I.,

et al. (2018). Improving language understanding by

generative pre-training.

Radford, A., Wu, J., Child, R., Luan, D., Amodei, D.,

Sutskever, I., et al. (2019). Language models are un-

supervised multitask learners. OpenAI blog, 1(8):9.

Trigui, O., Hadrich Belguith, L., Rosso, P., Ben Amor, H.,

and Gafsaoui, B. (2012). Arabic qa4mre at clef 2012:

Arabic question answering for machine reading eval-

uation.

WEBIST 2024 - 20th International Conference on Web Information Systems and Technologies

118