Madlyn-Manneh/NLP-Final-Project

Cross-lingual sentiment analysis for Krio using few-shot learning, multilingual transformers (mBERT and XLM-R), and balanced sampling techniques for low-resource NLP. # Zero-Shot and Few-Shot Cross-Lingual Sentiment Analysis from English to Krio ## Overview This project explores cross-lingual sentiment analysis for Krio, a low-resource English-lexified creole spoken in Sierra Leone. The study compares multilingual transformer models: - mBERT (`bert-base-multilingual-cased`) - XLM-R (`xlm-roberta-base`) for ternary sentiment classification: - Negative - Neutral - Positive The project investigates: - Zero-shot learning - Few-shot learning (10-shot and 50-shot) - Balanced few-shot training - Cross-lingual transfer from English to Krio --- ## Key Features - Few-shot learning for low-resource NLP - Cross-lingual sentiment transfer - mBERT vs XLM-R comparison - Balanced sampling experiments - TF-IDF + Logistic Regression baseline - Error analysis and confusion matrices --- ## Dataset ### English Dataset - `dair-ai/emotion` from Hugging Face - Mapped into: - Negative - Neutral - Positive ### Krio Dataset - 200 manually annotated Krio sentences - Sampled from `Jaward/Krio-Corpus` - Used for few-shot training and evaluation --- ## Models Used ### Baseline - TF-IDF Vectorizer - Logistic Regression ### Transformer Models - mBERT (`bert-base-multilingual-cased`) - XLM-R (`xlm-roberta-base`) --- ## Experimental Setup The project evaluates: | Setting | Description | |---|---| | Zero-shot | Train on English only | | 10-shot | Fine-tune with 10 Krio examples | | 50-shot | Fine-tune with 50 Krio examples | | Balanced 50-shot | Equal class sampling | --- ## Results | Experiment | mBERT | XLM-R | |---|---|---| | Zero-shot | 0.200 / 0.168 | 0.180 / 0.169 | | 10-shot | 0.573 / 0.263 | 0.227 / 0.123 | | 50-shot | 0.580 / 0.245 | 0.580 / 0.245 | | Balanced 50-shot | 0.453 / 0.324 | — | Format: - Accuracy / Macro F1 ### Key Findings - Few-shot learning significantly improved performance. - mBERT slightly outperformed XLM-R in this low-resource setting. - Balanced sampling improved Macro F1 and reduced Neutral-class bias. …