TaylorSergel/NLP_Group_6_Project

Multilingual emotion analysis across African languages using transfer learning and data augmentation on the BRIGHTER + EthioEmo datasets. # COS 760 — Emotion Analysis using BRIGHTER + EthioEmo Datasets **Group 5 | Alisha Perumal, Junior Motsepe, Taylor Sergel** Multilabel emotion classification across Afrikaans, English, isiZulu, Sesotho, and Setswana using transfer learning (XLM-RoBERTa, AfriBERTa) and data augmentation (back-translation, LLM annotation). --- ## Contents of the Zip File ``` NLP_Group_5_COS760/ ├── README.md ← This file ├── requirements.txt ← Python dependencies ├── notebooks/ │ ├── phase1_preprocessing.py ← Dataset acquisition and preprocessing │ ├── phase2_baseline.py ← TF-IDF + Logistic Regression + SVM baselines │ ├── phase2_visualisations.py ← Baseline result visualisations │ ├── phase3_ws1_xlmroberta.py ← XLM-RoBERTa LoRA fine-tuning │ ├── phase3_ws2_afriberta.py ← AfriBERTa LoRA fine-tuning │ ├── phase3_ws3_backtranslation.py ← NLLB-200 back-translation augmentation │ ├── phase3_ws4_llm_annotation.py ← Claude LLM-based emotion annotation │ ├── phase3_transfer_and_results.py ← Cross-lingual transfer + results aggregation │ └── phase4_error_analysis.py ← Error analysis and visualisations ├── results/ │ ├── phase2_baselines/ ← Baseline result CSVs │ └── phase2_visualisations/ ← Baseline charts and graphs └── models/ └── (empty — model checkpoints saved to Google Drive during training) ``` > **Note on data and model checkpoints:** All processed datasets and trained model checkpoints are stored on Google Drive due to file size constraints (datasets: ~500MB; model checkpoints: ~3GB). See the Data Information section below for download and setup instructions. --- ## Project Overview This project investigates multilabel emotion classification across five South African languages using the BRIGHTER + EthioEmo dataset. The pipeline covers five phases: | Phase | Description | |---|---|---| | Phase 1 | Dataset acquisition and preprocessing | | Phase 2 …