angiexbreedt/COS760-NLP-Project

Repository for our COS760 Natural Language Processing semester project (Semantic Relatedness Using SemRel Datasets - Evaluating Cross-Lingual Transfer and Data Augmentation for Semantic Relatedness in Afrikaans, Hausa and Kinyarwanda) # COS760 Group 16 - Semantic Relatedness Using SemRel Datasets **Evaluating Cross-Lingual Transfer and Data Augmentation for Semantic Relatedness in Afrikaans, Hausa and Kinyarwanda** ### 👥 Team Members - Angelique Breedt (u23542838) - Christopher Yoko (u22857941) - Resego Morei (u20570326) --- ## 📌 Project Overview Semantic relatedness measures how meaningfully connected two pieces of text are. This project investigates how well NLP models capture semantic relatedness in low-resource African languages. We focus on: - Afrikaans - Hausa - Kinyarwanda Using the **SemRel dataset**, we evaluate: - Cross-lingual transfer (train on English → test on African languages) - Multilingual vs African-specific models - Impact of data augmentation --- ## 🎯 Research Questions 1. How well do models trained on English transfer to African languages? 2. How do multilingual models compare to African-focused models? 3. Does data augmentation improve performance in low-resource settings? --- ## 🧠 Methods - Sentence embeddings (Sentence-BERT, XLM-R, AfriBERTa) - Cosine similarity for semantic relatedness - Fine-tuning on English and multilingual data - Data augmentation (back-translation, paraphrasing) --- ## 📊 Evaluation Metrics - Spearman Correlation (primary) - Mean Squared Error (MSE) --- ## Contents of the Zip File | File / Directory | Description | |---|---| | `README.md` | This file | | `requirements.txt` | Python package dependencies | | `Group16_u23542838.pdf` | Final project report | | `notebooks/01_eda.ipynb` | Exploratory data analysis - score distributions, sentence lengths, dataset statistics | | `notebooks/02_baseline.ipynb` | Cosine similarity baseline - runs on all 4 languages locally | | `notebooks/03-xlmr-finetuning.ipynb` | XLM-R fine-tuning and zero-shot evaluation - run on Kaggle GPU | | `notebooks/04-afriberta.ipynb` | AfriBERTa fine-tuning and zero-shot evaluation - run on Kaggle GPU | | `notebooks/05-augmentation.ipynb` | Back-translation augmentation …