Comparative Analysis of BiLSTM and AfriBERTa for Fake News Detection in Low-Resource Amharic Language
# Comparative Analysis of BiLSTM and AfriBERTa for Fake News Detection in Low-Resource Amharic Language
This repository contains the implementation, analysis, and documentation for the project **“Comparative Analysis of BiLSTM and AfriBERTa for Fake News Detection in Low-Resource Amharic Language.”**
The work compares a tuned BiLSTM architecture with the transformer-based AfriBERTa model for binary fake news detection in the Amharic language.
---
## 1. Project Overview
### Objective
To develop and evaluate deep learning models capable of detecting fake news in the Amharic language, addressing the lack of robust NLP solutions for low-resource linguistic settings.
### Models Examined
- **BiLSTM (Optimized Configuration)**
Custom architecture tuned for enhanced generalization and reduced overfitting.
- **AfriBERTa**
Transformer-based model adapted from XLM-RoBERTa and fine-tuned for Amharic fake news classification.
### Core Techniques
- Text preprocessing: normalization, cleaning, tokenization, and padding (308-token sequence length)
- Dataset balancing via SMOTE
- Training stabilization using early stopping, learning rate scheduling, and L2 regularization
- Evaluation metrics: Accuracy, Precision, Recall, and F1-Score
---
## 2. Dataset
**Source**
Hailu, M. (2024). *Amharic Fake News Detection Dataset.*
### Dataset Statistics
- Total samples: 8,630
- Fake: 4,185
- Real: 4,445
- Train/Validation/Test split: 70% / 15% / 15%
- Padding length: 308 tokens (95th percentile coverage)
---
## 3. Repository Structure
| File | Description |
|------|-------------|
| `Amharic_FakeNews_BiLSTM_AfriBERTa.ipynb` | Implementation of both models |
| `REPORT_Comparative_Analysis.pdf` | Complete technical research report |
| `PPT_Comparative_Analysis.pdf` | Presentation slides |
| `README.md` | Documentation file |
---
## 4. Model Performance
| Model | Accuracy | Precision | Recall | F1-Score | Train–Val Gap |
|-------|----------|-----------|--------|----------|----------- …