KatlehoMoeletsi/-Swahili-Sentiment-Analysis-using-Transformer-Models

# Swahili Sentiment Analysis using Transformer Models > **COS 760 – Natural Language Processing | University of Pretoria** A comparative study of classical machine learning and transformer-based approaches for sentiment analysis of Swahili tweets, using the AfriSenti dataset. --- ## Table of Contents - Project Overview - Research Questions - Results Summary - Repository Structure - Getting Started - Prerequisites - Installation - Data Setup - Running the Pipeline - Step 1: Preprocessing - Step 2: Baseline Models - Step 3: Transformer Models - Step 4: Data Augmentation (optional) - Step 5: Evaluation & Comparison - Notebooks - Testing - Key Findings - Responsible NLP - References --- ## Project Overview Sentiment analysis for African languages remains severely underexplored in NLP research. This project targets **Swahili** — spoken by ~200 million people across East Africa — and evaluates: | Model | Type | |---|---| | Naive Bayes (TF-IDF) | Classical baseline | | Logistic Regression (TF-IDF) | Classical baseline | | Linear SVM (TF-IDF) | Classical baseline | | XLM-RoBERTa (fine-tuned) | Transformer | | AfriBERTa (fine-tuned) | Transformer | We also examine the effect of SentencePiece subword tokenisation and back-translation data augmentation. --- ## Research Questions 1. **RQ1** — How do transformer-based models compare to classical baselines for Swahili sentiment analysis? 2. **RQ2** — What is the impact of subword tokenisation on model performance? 3. **RQ3** — Can back-translation augmentation improve classification results? 4. **RQ4** — What systematic error patterns exist, and how do they relate to code-switching? --- ## Results Summary | Model | Accuracy | F1 (Weighted) | Precision | Recall | |---|---|---|---|---| | Naive Bayes | 0.5604 | 0.5014 | 0.4705 | 0.5604 | | Logistic Regression | 0.6154 | 0.5159 | 0.5608 | 0.6154 | | Linear SVM | 0.5714 | 0.5142 | 0.4802 | 0.5714 | | XLM-RoBERTa | 0.5934 | 0.4420 | 0.3521 | 0.5934 | | AfriBERTa | 0.6 …