Multi-Domain Aspect-Based Sentiment Analysis Using Transformer-Based Models On Arabic Corpus /Yomna Eid Mohamed Rizk

Supervisor: Walaa Medhat

Thesis (M.A.)—Nile University, Egypt, 2025 .

"Includes bibliographical references"

Contents:
Contents
Page
CERTIFICATION OF APPROVAL ii
COPYRIGHT iii
ACKNOWLEDGEMENTS iv
ABSTRACT v
Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
LIST OF FIGURES xi
LIST OF TABLES xii
LIST OF EQUATIONS xiii
Chapters:
1. Introduction 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Thesis Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Organization of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2. Background and Concepts 5
2.1 Sentiment Analysis (SA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Aspect-Based Sentiment Analysis (ABSA) Task . . . . . . . . . . . . . . . . . . . . . 6
2.2.1 ABSA in the Arabic Language . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
vii
2.2.2 Models for ABSA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Multi-domain ABSA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3. Background and Literature Review 15
3.1 Existing ABSA Models in English . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2 Arabic Single-Domain ABSA Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.3 Arabic Single-Domain ABSA Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.4 Multi-domain ABSA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.4.1 Methodologies of Multi-Domain ABSA . . . . . . . . . . . . . . . . . . . . . . 23
3.4.2 Models for Multi-Domain ABSA . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4. Multi-Domain Aspect-Based Sentiment Analysis Proposed Model 26
4.1 A-MASA: Arabic Multi-Domain Aspect-Based Sentiment Analysis Datasets . . . . . . 27
4.1.1 Dataset Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.1.2 Dataset Cleaning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.1.3 Dataset Annotation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.1.4 Dataset Preparation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2 ABSA Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5. Experimental Evaluation 40
5.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.2 A-MASA Evaluation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.3 Muli-Domain ABSA Models Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.3.1 Muli-Domain Aspects Extraction Results . . . . . . . . . . . . . . . . . . . . . 45
5.3.2 Muli-Domain Aspects Sentiment Classification Results . . . . . . . . . . . . . 48
5.4 Other Models Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.4.1 Comparison Between Our Model and Arabic Single-domain AE Model . . . . . 49
5.4.2 Comparison Between Multi-domain Arabic ABSA Model and LLMs-based Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
6. Conclusions and Future Work 52
viii
6.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
6.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
Bibliography 60
Appendices:
A. Publications

Abstract:
Aspect-Based Sentiment Analysis (ABSA) is a fine-grained Sentiment Analysis (SA) task that
identifies sentiments expressed towards specific aspects within a sentence. Despite the increasing use
of Semitic languages like Arabic in Natural Language Processing (NLP), there remains a scarcity of
resources and datasets for Arabic ABSA. To address these challenges, we introduce a new manually
annotated multi-domain dataset, A-MASA (Arabic Multi-domain Aspect-based Sentiment Analysis), comprising 6,500 records from different domains and dialects. Additionally, a generalized model
for Arabic multi-domain ABSA was developed and trained on both A-MASA and existing datasets,
covering a total of nine domains. The proposed model integrates three state-of-the-art pre-trained
models, each fine-tuned with additional layers, and combines their outputs using a weighted sum ensemble technique. Comprehensive experiments were conducted to evaluate the model’s performance.
For the Aspect Extraction (AE) task, the QARiB model achieved the highest F1-score of 88.7% on
the hotels domain, while the ensemble model reached the highest F1-score of 88.9% on the products
domain. For aspect sentiment classification, the ensemble model recorded an F1-score of 87.5% on
the hotels domain. The ensemble model demonstrated superior generalization, outperforming existing single-domain models on unseen data, achieving 37% improvement in the books domain and
14% in the hotels domain for the AE task. Additionally, the model surpassed two Large Language
Models (LLMs), achieving an F1-score of 88.9% in the products domain 49% higher than JAIS and
31% higher than Gemini.
keywords
Arabic Aspect-Based Sentiment Analysis (ABSA), Multi-domain Sentiment Analysis, Natural Language Processing (NLP), Ensemble Learning Models, Arabic Datasets.

Text in English, abstracts in English and Arabic