Unis237/Cameroon-pidgin-news-classification

# News Topic Classification in Cameroonian Pidgin English ## Overview This project investigates the effectiveness of classical machine learning models and a fine-tuned multilingual transformer model for news topic classification in **Cameroonian Pidgin English**. The work was conducted as part of the **Natural Language Processing and Large Language Models course at the African Institute for Mathematical Sciences (AIMS Cameroon)**. The objective is to evaluate whether multilingual transformer models such as **AfroXLMR** can achieve strong performance on a low-resource African language classification task. --- ## Research Objective The study aims to: - Construct a dataset of Cameroonian Pidgin English news headlines. - Train classical machine learning baseline models. - Fine-tune the **AfroXLMR multilingual transformer model** for topic classification. - Compare the performance of classical and transformer-based approaches. --- ## Dataset ### Data Source The dataset consists of news headlines generated in **Cameroonian Pidgin English**. The headlines were generated using **Claude AI** based on prompts designed to produce examples corresponding to six news categories. The generated headlines were subsequently **reviewed and corrected by a fluent speaker of Cameroonian Pidgin English** to ensure linguistic accuracy. ### Topic Categories The dataset contains six topic labels: - Business - Entertainment - Health - Politics - Sports - Technology ### Dataset Distribution (Training Set) | Category | Samples | |--------|--------| | Politics | 163 | | Business | 151 | | Sports | 149 | | Health | 149 | | Technology | 149 | | Entertainment | 139 | ### Dataset Split | Dataset | Samples | |-------|-------| | Training | 900 | | Development | 100 | | Test | 150 | --- ## Methods ### Classical Machine Learning Models The following baseline models were trained using a **bag-of-words representation** produced by `CountVectorizer`: - Multinomial Naive Bayes - Multi-La …