Translate-Train (or Translate-Then-Train) is a cross-lingual transfer strategy where training data in a source language (e.g., English) is translated into the target language (e.g., Swahili) using Machine Translation, and the model is then fine-tuned on this synthesized data — converting a zero-shot problem into a supervised problem using synthetic data.

Mechanism

• Source: English labeled dataset (e.g., SQuAD).
• Translation: Use Google Translate/NLLB to translate SQuAD to Swahili.
• Alignment: Project labels (indices for spans) to the new text — the hardest part (requires alignment tools like Awesome-Align).
• Training: Fine-tune the model on the translated Swahili data.

Why It Matters

• Performance: Often outperforms Zero-Shot Transfer (fine-tune En, test Swahili) because the model sees actual Swahili tokens during training.
• Noise Tolerant: Deep learning models are surprisingly robust to translation noise (bad grammar in training data).
• Baseline: The standard baseline to beat in all cross-lingual papers.

Translate-Train is synthetic supervision — using machine translation to generate training data for languages that have none.