ibrahimbukhari1998/-Zero-Shot-for-Under-Resourced-Language

Zero-Shot POS Tagging for Under-Resourced Languages # Cross-Lingual POS Tagging: XLM-R vs. Glot500 ## Project Goal Compare the performance of XLM-R and Glot500 when fine-tuned for POS tagging on a better-resourced language and then applied directly to a low-resource language without further training. Analyze the impact of subword tokenization on cross-lingual transfer. ## Results ### Fragmentation Rate |High-Resource Language| XLMR Fragment Rate| Glot500 Fragment Rate| |--------------|----------|----------| | English|1.30 | 1.19| | French| 1.44|1.32 | | Standard Arabic| 1.0| 1.0| | Russian| 1.67|1.49 | |Low-Resource Language| XLMR Fragment Rate| Glot500 Fragment Rate| |--------------|----------|----------| | Wolof| 1.81 |1.38 | | Catalan|1.41 |1.30 | | Urdu| 1.31| 1.28| | Ukranian| 1.74| 1.57| ### Cross-Lingual Transfer Performance: Mono-Lingual Fine-Tuning Dataset | Language Pair | Model | Accuracy | F1 Score | |--------------|---------|----------|-----------| | English → Wolof | XLM-R | 37% | 34% | | | Glot500 | 47% | 46% | | Standard Arabic → Urdu | XLM-R | 17% | 11% | | | Glot500 | 24% | 10% | | French → Catalan | XLM-R | 47% | 47% | | | Glot500 | 68% | 68% | | Russian → Ukrainian | XLM-R | 54% | 53% | | | Glot500 | 72% | 69% | | Welsh → Irish | XLM-R | 47% | 42% | | | Glot500 | 33% | 19% | ### Cross-Lingual Transfer Performance: Mono-Lingual Fine-Tuning Dataset with Noise Injection | Language Pair | Model | 25% Noise | | 75% Noise | | |--------------|--------|------------|------------|------------|------------| | | | Accuracy | F1 Score | Accuracy | F1 Score | | English → Wolof | XLM-R | 23% | 20% | 22% | 19% | | | Glot500 | 42% | 41% |39% | 38% | | Standard Arabic → Urdu | XLM-R | 24% | 11% | 23% | 11% | | | Glot500 | 24% | 11% | 24% | 11% | | French → Catalan | XLM-R | 46% | 46% | 46% | 45% | | | Glot500 | 68% | 67% | 67% | 66% | | Russian → Ukrainian | XLM-R | 54% | 53% | 51% | 50% | | | Glot500 | 66% | 61% | 59% …