Alhibb/COEN541-_Natural_Language_Processing

Development of a Hausa-language news topic classifier using transfer learning techniques for low-resource scenarios. The application classifies VOA Hausa news headlines into 5 categories: Nigeria, Africa, World, Health, and Politics. # DEPARTMENT OF COMPUTER ENGINEERING. AHMADU BELLO UNIVERSITY, ZARIA FIRST SEMESTER EXAMINATION 2024/2025 ACADEMIC SESSION COEN541: Natural Language Processing --- ## Hausa News Topic Classification Application **Group 10 Project Submission** ### Group Members | Name | Matric Number | Role | |-------------------------------|----------------|---------------| | Ibrahim Rabiu | U19CO2013 | Group Leader | | Ibrahim Abduljabbar Hamid | U19CO2017 | Member | | Aminu Muhammad Baba | U18CO1035 | Member | | Arop Dominic Arop | U18CO1086 | Member | | Ya'u Musbahu Usman | U19CO2018 | Member | --- ## Table of Contents 1. Problem Statement 2. Objectives 3. Dataset 4. Methodology 5. Implementation 6. How to Run 7. Results 8. Challenges 9. Report Structure 10. References --- ## Problem Statement Development of a Hausa-language news topic classifier using transfer learning techniques for low-resource scenarios. The application classifies VOA Hausa news headlines into 5 categories: Nigeria, Africa, World, Health, and Politics. --- ## Objectives 1. Implement fine-tuning of pre-trained XLM-RoBERTa model 2. Address challenges in low-resource language processing 3. Develop interactive Gradio web interface 4. Achieve state-of-the-art performance on Hausa text classification 5. Demonstrate practical NLP application deployment --- ## Dataset **hausa_voa_topics** (From Hugging Face Datasets) - Source: `UdS-LSV/hausa_voa_topics` - Structure: - Total samples: 2,917 - Train: 2,045 samples - Validation: 290 samples - Test: 582 samples - Label Distribution (Training Set): | Label | Category | Count | |-------|----------|-------| | 4 | Politics | 563 | | 2 | World | 449 | | 1 | Africa | 434 | | 0 | Nigeria | 315 | | 3 | Health | 284 | --- ## Methodology 1. **Base Model**: XLM-RoBERTa-base (Uncased) 2. **Toke …