| Titre : | Multi-Task Learning for Mental Health Disorder Classification from User-Written Text |
| Auteurs : | TAHI Boumediene Amir, Auteur ; Kadari Rekia, Directeur de thèse ; YAHLALI Mebarka, Directeur de thèse |
| Type de document : | texte manuscrit |
| Editeur : | جامعة سعيدة د.مولاي الطاهر -كلية الرياضيات والاعلام الالي والاتصالات السلكية واللاسلكية -قسم الاعلام الالي, 2025/2026 |
| Format : | 82ص |
| Accompagnement : | CD |
| Note générale : |
General Introduction
Context and Motivation Mental health disorders such as depression, anxiety, bipolar disorder, and suicidal ideation represent a major global public health challenge, affecting hundreds of millions of indi- viduals worldwide. Traditional diagnostic pathways rely heavily on clinical interviews and self-reported questionnaires, which are inherently limited by delayed help-seeking be- haviour, social stigma, and restricted access to mental health professionals—particularly in low-resource settings. In parallel, the unprecedented growth of social media platforms and online communities has produced vast amounts of user-generated text that implicitly encodes emotional states, cognitive patterns, and behavioural cues. This convergence has opened a promising avenue for computational approaches capable of detecting early warn- ing signs of mental health disorders directly from naturally occurring written language, offering the potential for scalable, non-intrusive, and timely screening tools. Problem Statement Despite substantial progress in natural language processing (NLP), automatic mental health disorder classification from text remains a challenging problem for several reasons. First, the linguistic manifestations of different disorders frequently overlap—symptoms of depression and anxiety, for instance, share many surface-level features—making fine- grained multi-class classification considerably harder than binary detection. Second, pub- licly available datasets are often class-imbalanced, with critical minority classes (e.g., sui- cidal ideation, personality disorders) underrepresented relative to majority classes. Third, single-task classification models, trained solely to predict a disorder label, may fail to cap- ture the broader emotional and contextual signals that a human clinician would naturally consider when forming a diagnostic impression. This raises the central research question addressed in this thesis: can a multi-task learning framework, jointly optimising mental health disorder classification alongside a complementary auxiliary task such as sentiment analysis, yield richer textual representations and improved classification performance com- pared to single-task baselines? |
| Langues: | Français |
| Index. décimale : | BUC-M 008512 |
| Catégories : | |
| Mots-clés: | cessing, Transformer Models, DeBERTa, DistilBERT, Sentiment Analysis, Computational Mental Health, Deep Learning. |
| Résumé : |
Mental health disorders such as depression, anxiety, bipolar disorder, stress, personality
disorders, and suicidal ideation affect millions of people worldwide and remain difficult to identify at an early stage. Recent advances in Natural Language Processing (NLP) have enabled the automatic analysis of user-written text for mental health assessment. This thesis investigates the use of Multi-Task Learning (MTL) for mental health disorder classification from user-written text, where disorder prediction is learned jointly with sen- timent analysis as an auxiliary task. Several deep learning architectures were evaluated, including attention-based models, transformer-based models, and hybrid approaches, us- ing two publicly available mental health datasets. Experimental results show that trans- former architectures significantly outperform traditional approaches, with DeBERTa-base achieving the best performance, reaching 94.06% accuracy and 93.09% macro F1-score. The study also examines the impact of multi-task learning, model design choices, and training strategies on classification performance. The findings confirm the effectiveness of transformer-based models for mental health text classification and contribute to the development of accurate and scalable computational mental health screening systems. Keywords: Mental Health Classification, Multi-Task Learning, Natural Language Pro- |
| Note de contenu : |
Contents
Acknowledgements i Abstract 1 General Introduction 1 1 State of the Art: Multi-Task Learning for Mental Health Disorder Clas- sification from User-Written Text 4 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.2 Foundational Concepts in Multi-Task Learning for Mental Health Text Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 1.2.1 Why Multi-Task Learning for Mental Health? . . . . . . . . . . . . 5 1.2.2 Hard and Soft Parameter Sharing Approaches . . . . . . . . . . . . 6 1.3 Architectural Innovations (2020–2025) . . . . . . . . . . . . . . . . . . . . 6 1.3.1 Hierarchical and Theory-Driven Multi-Task Architectures . . . . . 6 1.3.2 Attention Mechanisms and Multi-Task Fusion . . . . . . . . . . . . 7 1.3.3 Prompt-Based and Fine-Grained Multi-Task Learning . . . . . . . 7 1.3.4 Cross-Lingual and Language-Agnostic MTL . . . . . . . . . . . . . 7 1.4 Auxiliary Tasks in Mental Health MTL . . . . . . . . . . . . . . . . . . . . 8 1.4.1 Emotion and Sentiment Analysis . . . . . . . . . . . . . . . . . . . 8 1.4.2 Cognitive Distortion Detection . . . . . . . . . . . . . . . . . . . . 8 1.4.3 Multi-Label and Comorbidity Modelling . . . . . . . . . . . . . . . 8 1.5 Datasets and Evaluation Frameworks . . . . . . . . . . . . . . . . . . . . . 9 1.5.1 Social Media Corpora . . . . . . . . . . . . . . . . . . . . . . . . . 9 1.5.2 Evaluation Metrics and Baselines . . . . . . . . . . . . . . . . . . . 9 1.6 Challenges and Open Problems . . . . . . . . . . . . . . . . . . . . . . . . 9 1.6.1 Task Interference and Negative Transfer . . . . . . . . . . . . . . . 9 1.6.2 Privacy and Ethical Considerations . . . . . . . . . . . . . . . . . . 10 1.6.3 Data Scarcity and Annotation Quality . . . . . . . . . . . . . . . . 10 1.6.4 Temporal Dynamics and Longitudinal Modelling . . . . . . . . . . 10 ii 1.7 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 1.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2 Background 12 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 2.2 Artificial Intelligence in Healthcare . . . . . . . . . . . . . . . . . . . . . . 13 2.2.1 Overview of AI Applications in Clinical Settings . . . . . . . . . . . 13 2.2.2 Natural Language Processing in Medical Analysis . . . . . . . . . . 14 2.2.3 Ethical, Privacy, and Interpretability Considerations . . . . . . . . 14 2.3 Mental Health Disorders . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 2.3.1 Anxiety Disorders . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 2.3.2 Depression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 2.3.3 Bipolar Disorder . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 2.3.4 Attention-Deficit/Hyperactivity Disorder (ADHD) . . . . . . . . . 16 2.3.5 Post-Traumatic Stress Disorder (PTSD) . . . . . . . . . . . . . . . 16 2.3.6 Personality Disorder . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.3.7 Stress . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.3.8 Suicidal Ideation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.3.9 Summary of Classification Challenges . . . . . . . . . . . . . . . . . 18 2.4 Natural Language Processing for Mental Health Analysis . . . . . . . . . . 18 2.4.1 Fundamentals of Text Classification . . . . . . . . . . . . . . . . . . 18 2.4.2 Sentiment Analysis and Emotion Detection . . . . . . . . . . . . . 20 2.4.3 Transformer-Based Language Models . . . . . . . . . . . . . . . . . 20 2.4.4 Challenges of Social Media Text for NLP . . . . . . . . . . . . . . . 21 2.5 Datasets Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 2.5.1 Reddit Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 2.5.2 MA Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.5.3 Comparative Dataset Analysis . . . . . . . . . . . . . . . . . . . . . 24 2.6 State of the Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 2.6.1 Transformer-Based Mental Health Classification . . . . . . . . . . . 25 2.6.2 Multi-Task Learning for NLP . . . . . . . . . . . . . . . . . . . . . 27 2.6.3 Social Media Mental Health Detection . . . . . . . . . . . . . . . . 28 2.6.4 Critical Comparative Analysis . . . . . . . . . . . . . . . . . . . . . 29 2.6.5 Identified Research Gaps . . . . . . . . . . . . . . . . . . . . . . . . 29 2.7 Problem Statement and Research Objectives . . . . . . . . . . . . . . . . . 30 2.7.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 2.7.2 Research Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . 31 iii 3 Methodology and Experimental Setup 33 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 3.2 Theoretical Foundations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 3.2.1 Attention Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . 34 3.2.2 Transformer Architecture . . . . . . . . . . . . . . . . . . . . . . . 35 3.2.3 Pre-Trained Language Models . . . . . . . . . . . . . . . . . . . . . 36 3.3 Data Pipeline and Preprocessing . . . . . . . . . . . . . . . . . . . . . . . 37 3.3.1 Dataset Description . . . . . . . . . . . . . . . . . . . . . . . . . . 37 3.3.2 Text Preprocessing Pipeline . . . . . . . . . . . . . . . . . . . . . . 38 3.3.3 Linguistic Feature Extraction . . . . . . . . . . . . . . . . . . . . . 39 3.3.4 Word Embedding Features . . . . . . . . . . . . . . . . . . . . . . . 40 3.3.5 Data Augmentation . . . . . . . . . . . . . . . . . . . . . . . . . . 41 3.3.6 Class Imbalance Handling . . . . . . . . . . . . . . . . . . . . . . . 41 3.4 Multi-Task Learning Framework . . . . . . . . . . . . . . . . . . . . . . . . 41 3.4.1 Theoretical Motivation . . . . . . . . . . . . . . . . . . . . . . . . . 41 3.4.2 Hard Parameter Sharing . . . . . . . . . . . . . . . . . . . . . . . . 42 3.4.3 Multi-Task Loss Formulation . . . . . . . . . . . . . . . . . . . . . 42 3.4.4 Sentiment Label Derivation . . . . . . . . . . . . . . . . . . . . . . 43 3.5 Model Architectures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 3.5.1 Family I: Scratch-Built Attention Models . . . . . . . . . . . . . . . 43 3.5.2 Family II: Fine-Tuned Transformer Models . . . . . . . . . . . . . . 45 3.5.3 Family III: Hybrid Feature-Fusion Model . . . . . . . . . . . . . . . 46 3.5.4 Multi-Seed Ensemble (Experiment 17) . . . . . . . . . . . . . . . . 47 3.6 Training Procedures and Optimisation . . . . . . . . . . . . . . . . . . . . 48 3.6.1 Optimiser Configuration . . . . . . . . . . . . . . . . . . . . . . . . 48 3.6.2 Learning Rate Scheduling . . . . . . . . . . . . . . . . . . . . . . . 48 3.6.3 Mixed-Precision Training . . . . . . . . . . . . . . . . . . . . . . . 48 3.6.4 Gradient Clipping . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 3.6.5 Regularisation Techniques . . . . . . . . . . . . . . . . . . . . . . . 48 3.6.6 Comprehensive Hyperparameter Summary . . . . . . . . . . . . . . 49 3.6.7 Hardware and Computational Environment . . . . . . . . . . . . . 49 3.7 Evaluation Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 3.7.1 Evaluation Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 3.7.2 Evaluation Procedure . . . . . . . . . . . . . . . . . . . . . . . . . 50 3.7.3 Visualisation and Reporting . . . . . . . . . . . . . . . . . . . . . . 51 4 Results and Discussion 52 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 4.2 Experimental Configurations . . . . . . . . . . . . . . . . . . . . . . . . . . 53 iv 4.2.1 Experiment Taxonomy . . . . . . . . . . . . . . . . . . . . . . . . . 53 4.2.2 Hyperparameter Configurations . . . . . . . . . . . . . . . . . . . . 53 4.3 Evaluation Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 4.3.1 Metric Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 4.3.2 Metric Relevance for Mental Health Classification . . . . . . . . . . 55 4.4 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 4.4.1 Family I: Scratch-Built Attention Models . . . . . . . . . . . . . . . 56 4.4.2 Family II: Single-Task Fine-Tuned Transformers (MA Dataset) . . 57 4.4.3 Family III: Multi-Task Learning Models (Reddit Dataset) . . . . . 58 4.4.4 Family IV: Hybrid Feature-Fusion Model (MA Dataset) . . . . . . 58 4.4.5 Family V: Multi-Seed Ensemble (MA Dataset) . . . . . . . . . . . . 59 4.5 Comparative Analysis with State of the Art . . . . . . . . . . . . . . . . . 60 4.5.1 Cross-Model Performance Ranking . . . . . . . . . . . . . . . . . . 60 4.5.2 Comparison with Published Approaches . . . . . . . . . . . . . . . 60 4.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 4.6.1 The Dominance of Pre-Trained Transformers . . . . . . . . . . . . 61 4.6.2 DeBERTa vs. DistilBERT: Architectural Insights . . . . . . . . . . 61 4.6.3 The Failure of Feature Fusion . . . . . . . . . . . . . . . . . . . . . 62 4.6.4 Multi-Task Learning: Promise and Limitations . . . . . . . . . . . 62 4.6.5 Class-Level Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 62 4.6.6 The Focal Loss Failure . . . . . . . . . . . . . . . . . . . . . . . . . 63 4.6.7 Computational Considerations . . . . . . . . . . . . . . . . . . . . . 63 4.7 Limitations and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . 63 4.7.1 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 4.7.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 General Conclusion 66 References 68 v |
Exemplaires
| Code-barres | Cote | Support | Localisation | Section | Disponibilité |
|---|---|---|---|---|---|
| aucun exemplaire |
Documents numériques (2)
BUC-M 008512 Adobe Acrobat PDF | BUC-M 008512 Adobe Acrobat PDF |
