|
| Titre : |
AI-Driven Fake News Detection: Applying Transformers /Large Language Models |
| Type de document : |
document électronique |
| Auteurs : |
Raid Nedjm Eddine Dekkar ; Mohamed El Amine Bouchareb, Auteur ; Nadjet Kamel, Directeur de thèse |
| Editeur : |
Setif:UFA |
| Année de publication : |
2025 |
| Importance : |
1 vol (84 f .) |
| Format : |
29 cm |
| Langues : |
Anglais (eng) |
| Catégories : |
Thèses & Mémoires:Informatique
|
| Mots-clés : |
Fake news detection
Algerian Dialect
Natural Language Processing
Transformers
Large Language Models
Social Media
Arabic NLP |
| Index. décimale : |
004 Informatique |
| Résumé : |
The proliferation of fake news across digital platforms has emerged as a major concern,
particularly when misinformation is expressed in low-resource dialects such as Algerian
Arabic. The Algerian dialect poses specific challenges for automatic detection due to its
rich linguistic variability, lack of standardized orthography, and limited availability of
natural language processing (NLP) tools tailored to it. This thesis explores the problem
of fake news detection in the Algerian dialect, not only in social media content but also in
other digital sources such as online articles and news websites. We investigate the effectiveness
of transformer-based models and large language models (LLMs) for this task, conducting
extensive experiments using a variety of pre-trained models including AraGPT2,
DziriBERT, CaMelBERT, LLaMa3, Qwen and others. The experimental findings
confirm that transformer and LLM-based approaches provide promising performance for
fake news detection in under-resourced languages, with AraGPT2 achieving the highest
accuracy in our evaluation benchmarks. Additionally, we propose a hybrid strategy
that combines transformer-based embeddings with LLM inference to better capture the
contextual and linguistic nuances of the dialect. |
| Note de contenu : |
Sommaire
1 General Introduction 1
Chapter 1. General Introduction 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 The Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.5 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 Background and Literature Review 5
Chapter 2. Background and Literature Review 5
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Understanding Fake News . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2.1 Concepts and Definition . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2.2 Fake News Categories . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2.3 Fake News Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Fake News and Social Networks . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3.2 Platforms of Social Networks . . . . . . . . . . . . . . . . . . . . . 11
2.3.3 Challenges of Social Networks . . . . . . . . . . . . . . . . . . . . . 12
2.4 Fake News Detection and NLP . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4.1 Natural Language Processing (NLP) . . . . . . . . . . . . . . . . . 13
2.4.2 Linguistic Component . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4.2.1 The Pre-Processing Phase: Transforming Text into Data . 14
2.4.2.2 Feature Extraction Techniques . . . . . . . . . . . . . . . 15
2.4.2.3 Learning Phase: From Data to Model . . . . . . . . . . . 16
2.5 Fake News Detection Approaches . . . . . . . . . . . . . . . . . . . . . . . 17
2.5.1 Machine Learning Approach . . . . . . . . . . . . . . . . . . . . . . 17
2.5.1.1 Machine Learning Categories: . . . . . . . . . . . . . . . . 17
2.5.2 Deep Learning Approach . . . . . . . . . . . . . . . . . . . . . . . . 18
2.5.2.1 Multi Layer Perceptron (MLP) . . . . . . . . . . . . . . . 19
2.5.2.2 Other Architectures: . . . . . . . . . . . . . . . . . . . . . 19
2.5.3 Transformers Approach . . . . . . . . . . . . . . . . . . . . . . . . 19
2.5.3.1 Transformer Architecture . . . . . . . . . . . . . . . . . . 20
2.5.3.2 Advantages of Transformer Models . . . . . . . . . . . . . 22
2.6 Transformer Based Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.6.1 Transformer-Based Models for Arabic and Algerian Dialect: . . . . 23
2.7 Large Language Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.7.1 From Pre-trained Language Models to LLMs . . . . . . . . . . . . 25
2.7.2 Zero-Shot, One-Shot, and Few-Shot Learning . . . . . . . . . . . . 25
2.8 Fake News Detection in Arabic . . . . . . . . . . . . . . . . . . . . . . . . 27
2.8.1 Arabic Language: . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.8.2 Categories of Arabic Language: . . . . . . . . . . . . . . . . . . . . 27
2.8.3 Arabic Language NLP Challenges: . . . . . . . . . . . . . . . . . . 28
2.9 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.9.1 ML Based Approach . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.9.2 DL Based Approach . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.9.3 Transformers / LLM Based Approach . . . . . . . . . . . . . . . . 30
2.10 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.10.1 Identified Research Gaps . . . . . . . . . . . . . . . . . . . . . . . . 32
3 Datasets Exploration 33
Chapter 3. Datasets Exploration 33
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 FASSILA Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2.1 Dataset Overview and Foundation . . . . . . . . . . . . . . . . . . 33
3.2.1.1 Corpus Origins and Development Context . . . . . . . . . 33
3.2.1.2 Functional Objectives and Target Applications . . . . . . 34
3.2.1.3 Linguistic Characteristics and Script Composition . . . . 34
3.2.2 Structural Composition \ Quantitative Analysis . . . . . . . . . . . 34
3.2.2.1 Corpus Architecture and Data Schema . . . . . . . . . . . 34
3.2.2.2 Source Distribution and Data Provenance . . . . . . . . . 35
3.2.2.3 Category Distribution and Domain Coverage . . . . . . . 36
3.2.2.4 Classification Distribution and Label Balance . . . . . . . 36
3.2.3 Data Preprocessing and Normalization . . . . . . . . . . . . . . . . 37
3.2.3.1 Collection Methodology . . . . . . . . . . . . . . . . . . . 37
3.2.3.2 Text Cleaning and Normalization . . . . . . . . . . . . . . 37
3.2.3.3 Controlled Augmentation via GPT-4 . . . . . . . . . . . . 37
3.2.3.4 Dialectal Geographic Distribution . . . . . . . . . . . . . 38
3.2.3.5 Word Clouds . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.2.4 Dataset Limitations and Constraints . . . . . . . . . . . . . . . . . 39
3.3 DZ-AFND Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.3.1 Dataset Origin and Motivation . . . . . . . . . . . . . . . . . . . . 39
3.3.2 Structural Composition and Label Distribution . . . . . . . . . . . 40
3.3.3 Translation Strategy and Gemini API Configuration . . . . . . . . 40
3.3.4 Classification Distribution and Label Balance . . . . . . . . . . . . 41
3.3.5 Prompting Technique and Constraints . . . . . . . . . . . . . . . . 41
3.3.6 Challenges and Limitations . . . . . . . . . . . . . . . . . . . . . . 42
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4 Experiments and Results 44
Chapter 4. Experiments and Results 44
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.2 Programming Language, Tools and Libraries . . . . . . . . . . . . . . . . . 44
4.2.1 Programming Language . . . . . . . . . . . . . . . . . . . . . . . . 44
4.2.2 Libraries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.2.3 Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.2.4 Evaluation Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.3 System Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.4 Data Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.4.1 Experimental Dataset . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.4.2 Data Cleaning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.4.2.1 Removing Stop Words . . . . . . . . . . . . . . . . . . . . 49
4.4.2.2 Removing Emojis, Special Caracters and Extra Spaces . . 50
4.4.2.3 Converting latin words to their Arabic alternative . . . . 50
4.4.3 Importing Dependencies . . . . . . . . . . . . . . . . . . . . . . . . 51
4.4.4 Loading Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.5 Fine-Tuning and Training Parameters . . . . . . . . . . . . . . . . . . . . . 53
4.5.1 Fine-Tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.5.2 Training Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.6 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.6.1 Phase 1: Baseline Experiments (No Fine-Tuning) . . . . . . . . . . 55
4.6.1.1 Experiment 1: Qwen2.5 (Zero-Shot) . . . . . . . . . . . . 55
4.6.4.6.1.3 Experiment 3: TinyLLaMA Enhanced (Prompt Engineering)
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.6.1.4 Experiment 4: DZiriBERT + Logistic Regression . . . . . 56
4.6.1.5 Experiment 5: DZiriBERT + MLP Classifier . . . . . . . 57
4.6.2 Phase 2: Fine-Tuned Transformers & LLMs . . . . . . . . . . . . . 58
4.6.2.1 Experiment 1: LLaMA-3 8B (Fine-Tuned) . . . . . . . . . 59
4.6.2.2 Experiment 2: AraElectra (Fine-Tuned) . . . . . . . . . . 59
4.6.2.3 Experiment 3: bert-base-multilingual-cased (Fine-Tuned) 60
4.6.2.4 Experiment 4: MARBERT (Fine-Tuned) . . . . . . . . . 60
4.6.2.5 Experiment 5: CAMeLBERT DA (Fine-Tuned) . . . . . . 60
4.6.2.6 Experiment 6: DZiriBERT (Fine-Tuned) . . . . . . . . . 60
4.6.2.7 Experiment 7: AraBERT-LoRA + Linear Classifier . . . . 61
4.6.2.8 Experiment 8: AraBERT-LoRA (Enhanced) . . . . . . . . 61
4.6.2.9 Experiment 9: AraGPT2-Base (Fine-Tuned) . . . . . . . 62
4.6.2.10 Experiment 10: AraGPT2-Medium (Fine-Tuned) . . . . . 62
4.6.2.11 Experiment 11: AraGPT2-Large (Fine-Tuned) . . . . . . 62
4.6.2.12 Experiment 12: AraGPT2-Mega (Fine-Tuned) . . . . . . 62
4.6.2.13 Additional Experiments on DZ-AFND . . . . . . . . . . . 63
4.6.3 Phase 3: Our AppHybrid (Embeddings + LLMs) . . . . . . . . . . 64
4.6.3.1 Experiment 1: TinyLLaMA + DZiriBERT Embeddings . 65
4.6.3.2 Experiment 2: LLaMA-3 8B + MARBERT Embeddings . 65
4.7 Discussion and Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.7.1 Achieving and Surpassing the Benchmark . . . . . . . . . . . . . . 70
4.7.2 Phase-Wise Insights . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.7.3 Model Trade-Offs and Practical Implications . . . . . . . . . . . . . 71
4.7.4 Final Comparison with the 2024 Benchmark . . . . . . . . . . . . . 71
4.8 Web App Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5 General Conclusion and Future Work 74
Chapter 5. General Conclusion and Future Work 74
5.1 General Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
|
| Côte titre : |
MAI/1018 |
AI-Driven Fake News Detection: Applying Transformers /Large Language Models [document électronique] / Raid Nedjm Eddine Dekkar ; Mohamed El Amine Bouchareb, Auteur ; Nadjet Kamel, Directeur de thèse . - [S.l.] : Setif:UFA, 2025 . - 1 vol (84 f .) ; 29 cm. Langues : Anglais ( eng)
| Catégories : |
Thèses & Mémoires:Informatique
|
| Mots-clés : |
Fake news detection
Algerian Dialect
Natural Language Processing
Transformers
Large Language Models
Social Media
Arabic NLP |
| Index. décimale : |
004 Informatique |
| Résumé : |
The proliferation of fake news across digital platforms has emerged as a major concern,
particularly when misinformation is expressed in low-resource dialects such as Algerian
Arabic. The Algerian dialect poses specific challenges for automatic detection due to its
rich linguistic variability, lack of standardized orthography, and limited availability of
natural language processing (NLP) tools tailored to it. This thesis explores the problem
of fake news detection in the Algerian dialect, not only in social media content but also in
other digital sources such as online articles and news websites. We investigate the effectiveness
of transformer-based models and large language models (LLMs) for this task, conducting
extensive experiments using a variety of pre-trained models including AraGPT2,
DziriBERT, CaMelBERT, LLaMa3, Qwen and others. The experimental findings
confirm that transformer and LLM-based approaches provide promising performance for
fake news detection in under-resourced languages, with AraGPT2 achieving the highest
accuracy in our evaluation benchmarks. Additionally, we propose a hybrid strategy
that combines transformer-based embeddings with LLM inference to better capture the
contextual and linguistic nuances of the dialect. |
| Note de contenu : |
Sommaire
1 General Introduction 1
Chapter 1. General Introduction 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 The Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.5 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 Background and Literature Review 5
Chapter 2. Background and Literature Review 5
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Understanding Fake News . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2.1 Concepts and Definition . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2.2 Fake News Categories . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2.3 Fake News Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Fake News and Social Networks . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3.2 Platforms of Social Networks . . . . . . . . . . . . . . . . . . . . . 11
2.3.3 Challenges of Social Networks . . . . . . . . . . . . . . . . . . . . . 12
2.4 Fake News Detection and NLP . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4.1 Natural Language Processing (NLP) . . . . . . . . . . . . . . . . . 13
2.4.2 Linguistic Component . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4.2.1 The Pre-Processing Phase: Transforming Text into Data . 14
2.4.2.2 Feature Extraction Techniques . . . . . . . . . . . . . . . 15
2.4.2.3 Learning Phase: From Data to Model . . . . . . . . . . . 16
2.5 Fake News Detection Approaches . . . . . . . . . . . . . . . . . . . . . . . 17
2.5.1 Machine Learning Approach . . . . . . . . . . . . . . . . . . . . . . 17
2.5.1.1 Machine Learning Categories: . . . . . . . . . . . . . . . . 17
2.5.2 Deep Learning Approach . . . . . . . . . . . . . . . . . . . . . . . . 18
2.5.2.1 Multi Layer Perceptron (MLP) . . . . . . . . . . . . . . . 19
2.5.2.2 Other Architectures: . . . . . . . . . . . . . . . . . . . . . 19
2.5.3 Transformers Approach . . . . . . . . . . . . . . . . . . . . . . . . 19
2.5.3.1 Transformer Architecture . . . . . . . . . . . . . . . . . . 20
2.5.3.2 Advantages of Transformer Models . . . . . . . . . . . . . 22
2.6 Transformer Based Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.6.1 Transformer-Based Models for Arabic and Algerian Dialect: . . . . 23
2.7 Large Language Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.7.1 From Pre-trained Language Models to LLMs . . . . . . . . . . . . 25
2.7.2 Zero-Shot, One-Shot, and Few-Shot Learning . . . . . . . . . . . . 25
2.8 Fake News Detection in Arabic . . . . . . . . . . . . . . . . . . . . . . . . 27
2.8.1 Arabic Language: . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.8.2 Categories of Arabic Language: . . . . . . . . . . . . . . . . . . . . 27
2.8.3 Arabic Language NLP Challenges: . . . . . . . . . . . . . . . . . . 28
2.9 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.9.1 ML Based Approach . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.9.2 DL Based Approach . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.9.3 Transformers / LLM Based Approach . . . . . . . . . . . . . . . . 30
2.10 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.10.1 Identified Research Gaps . . . . . . . . . . . . . . . . . . . . . . . . 32
3 Datasets Exploration 33
Chapter 3. Datasets Exploration 33
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 FASSILA Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2.1 Dataset Overview and Foundation . . . . . . . . . . . . . . . . . . 33
3.2.1.1 Corpus Origins and Development Context . . . . . . . . . 33
3.2.1.2 Functional Objectives and Target Applications . . . . . . 34
3.2.1.3 Linguistic Characteristics and Script Composition . . . . 34
3.2.2 Structural Composition \ Quantitative Analysis . . . . . . . . . . . 34
3.2.2.1 Corpus Architecture and Data Schema . . . . . . . . . . . 34
3.2.2.2 Source Distribution and Data Provenance . . . . . . . . . 35
3.2.2.3 Category Distribution and Domain Coverage . . . . . . . 36
3.2.2.4 Classification Distribution and Label Balance . . . . . . . 36
3.2.3 Data Preprocessing and Normalization . . . . . . . . . . . . . . . . 37
3.2.3.1 Collection Methodology . . . . . . . . . . . . . . . . . . . 37
3.2.3.2 Text Cleaning and Normalization . . . . . . . . . . . . . . 37
3.2.3.3 Controlled Augmentation via GPT-4 . . . . . . . . . . . . 37
3.2.3.4 Dialectal Geographic Distribution . . . . . . . . . . . . . 38
3.2.3.5 Word Clouds . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.2.4 Dataset Limitations and Constraints . . . . . . . . . . . . . . . . . 39
3.3 DZ-AFND Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.3.1 Dataset Origin and Motivation . . . . . . . . . . . . . . . . . . . . 39
3.3.2 Structural Composition and Label Distribution . . . . . . . . . . . 40
3.3.3 Translation Strategy and Gemini API Configuration . . . . . . . . 40
3.3.4 Classification Distribution and Label Balance . . . . . . . . . . . . 41
3.3.5 Prompting Technique and Constraints . . . . . . . . . . . . . . . . 41
3.3.6 Challenges and Limitations . . . . . . . . . . . . . . . . . . . . . . 42
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4 Experiments and Results 44
Chapter 4. Experiments and Results 44
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.2 Programming Language, Tools and Libraries . . . . . . . . . . . . . . . . . 44
4.2.1 Programming Language . . . . . . . . . . . . . . . . . . . . . . . . 44
4.2.2 Libraries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.2.3 Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.2.4 Evaluation Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.3 System Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.4 Data Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.4.1 Experimental Dataset . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.4.2 Data Cleaning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.4.2.1 Removing Stop Words . . . . . . . . . . . . . . . . . . . . 49
4.4.2.2 Removing Emojis, Special Caracters and Extra Spaces . . 50
4.4.2.3 Converting latin words to their Arabic alternative . . . . 50
4.4.3 Importing Dependencies . . . . . . . . . . . . . . . . . . . . . . . . 51
4.4.4 Loading Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.5 Fine-Tuning and Training Parameters . . . . . . . . . . . . . . . . . . . . . 53
4.5.1 Fine-Tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.5.2 Training Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.6 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.6.1 Phase 1: Baseline Experiments (No Fine-Tuning) . . . . . . . . . . 55
4.6.1.1 Experiment 1: Qwen2.5 (Zero-Shot) . . . . . . . . . . . . 55
4.6.4.6.1.3 Experiment 3: TinyLLaMA Enhanced (Prompt Engineering)
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.6.1.4 Experiment 4: DZiriBERT + Logistic Regression . . . . . 56
4.6.1.5 Experiment 5: DZiriBERT + MLP Classifier . . . . . . . 57
4.6.2 Phase 2: Fine-Tuned Transformers & LLMs . . . . . . . . . . . . . 58
4.6.2.1 Experiment 1: LLaMA-3 8B (Fine-Tuned) . . . . . . . . . 59
4.6.2.2 Experiment 2: AraElectra (Fine-Tuned) . . . . . . . . . . 59
4.6.2.3 Experiment 3: bert-base-multilingual-cased (Fine-Tuned) 60
4.6.2.4 Experiment 4: MARBERT (Fine-Tuned) . . . . . . . . . 60
4.6.2.5 Experiment 5: CAMeLBERT DA (Fine-Tuned) . . . . . . 60
4.6.2.6 Experiment 6: DZiriBERT (Fine-Tuned) . . . . . . . . . 60
4.6.2.7 Experiment 7: AraBERT-LoRA + Linear Classifier . . . . 61
4.6.2.8 Experiment 8: AraBERT-LoRA (Enhanced) . . . . . . . . 61
4.6.2.9 Experiment 9: AraGPT2-Base (Fine-Tuned) . . . . . . . 62
4.6.2.10 Experiment 10: AraGPT2-Medium (Fine-Tuned) . . . . . 62
4.6.2.11 Experiment 11: AraGPT2-Large (Fine-Tuned) . . . . . . 62
4.6.2.12 Experiment 12: AraGPT2-Mega (Fine-Tuned) . . . . . . 62
4.6.2.13 Additional Experiments on DZ-AFND . . . . . . . . . . . 63
4.6.3 Phase 3: Our AppHybrid (Embeddings + LLMs) . . . . . . . . . . 64
4.6.3.1 Experiment 1: TinyLLaMA + DZiriBERT Embeddings . 65
4.6.3.2 Experiment 2: LLaMA-3 8B + MARBERT Embeddings . 65
4.7 Discussion and Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.7.1 Achieving and Surpassing the Benchmark . . . . . . . . . . . . . . 70
4.7.2 Phase-Wise Insights . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.7.3 Model Trade-Offs and Practical Implications . . . . . . . . . . . . . 71
4.7.4 Final Comparison with the 2024 Benchmark . . . . . . . . . . . . . 71
4.8 Web App Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5 General Conclusion and Future Work 74
Chapter 5. General Conclusion and Future Work 74
5.1 General Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
|
| Côte titre : |
MAI/1018 |
|
AI-Driven Fake News Detection: Applying Transformers /Large Language Models
