arpit-shukla1/Swahili-News-Classification

# Swahili News Classification ## Overview This project focuses on **Swahili News Classification** using **Natural Language Processing (NLP)** techniques. The objective is to develop a machine learning model that classifies Swahili-language news articles into predefined categories. ## Dataset - **Source:** Zindi Africa Swahili News Classification Competition - **Categories:** - **Biashara** (Business) - **Burudani** (Entertainment) - **Kimataifa** (International News) - **Kitaifa** (National News) - **Michezo** (Sports) - **Challenges:** - **Severe class imbalance** (e.g., Burudani has only 2 samples) - **Overlapping topics** between categories - **Limited labeled Swahili NLP datasets** ## Preprocessing - **Data Cleaning:** Removed special characters and redundant spaces - **Tokenization:** Used Swahili-specific tokenizer - **Truncation & Padding:** Standardized input length to 512 tokens - **Stopword Removal:** Removed frequent but non-informative words - **Label Encoding:** Converted categorical labels into numerical format ## Model Selection - **Proposed Model:** - **RoBERTa-based model** fine-tuned for Swahili - **Pre-trained model:** `benjamin/roberta-base-wechsel-swahili` - **Final Accuracy:** **91.4%** ## Methodology - **Train-Validation Split:** 80% Training, 20% Validation - **Loss Function:** Cross-Entropy Loss (weighted for class imbalance) - **Optimizer:** AdamW - **Hyperparameters:** - **Batch Size:** 8 - **Learning Rate:** 2e-5 - **Epochs:** 10 ## Results | Metric | Score | |---------|------| | **Accuracy** | 91.40% | | **Precision (Weighted)** | 91.55% | | **Recall (Weighted)** | 91.40% | | **F1 Score (Weighted)** | 91.41% | ### **Per-Class Performance** | Class | Precision | Recall | F1 Score | |-------|-----------|-----------|-----------| | **Biashara** | 92.55% | 88.39% | 90.42% | | **Burudani** | 50.00% | 100.00% | 66.67% | | **Kimataifa** | 75.00% | 54.55% | 63.16% | | **Kitaifa** | 87.11% | 91.71% | 89.35% | | **Michezo** | 96.17% | 94.18 …