Cross-Lingual Transfer is the ability of a model trained on a task in a source language (e.g., English) to perform the same task in a target language (e.g., Japanese) without seeing any labeled training data in the target language — a capability emerging from multilingual pre-training.

Scenario

• Train: Fine-tune mBERT on SQuAD (English QA dataset).
• Test: Evaluate the model on a Japanese QA dataset.
• Result: The model performs surprisingly well, implying it learned "Question Answering" abstractly, independent of language.

Mechanisms

• Zero-Shot Transfer: No target language data used.
• Few-Shot Transfer: A few examples in target language provided.
• Alignment: Pre-training aligns embeddings so "cat" (En) and "gato" (Es) are close in vector space.

Why It Matters

• Global Scaling: Build an app for 100 languages while only labeling data for one.
• Equity: Brings state-of-the-art AI capabilities to languages with little labeled data.

Cross-Lingual Transfer is learn once, apply everywhere — leveraging high-resource language data to solve problems in low-resource languages.