zero-shot cross-lingual transfer

**Zero-Shot Cross-Lingual Transfer** is the **extreme case of cross-lingual transfer where the model is fine-tuned ONLY on the source language and then immediately evaluated on the target language with ZERO additional training or examples**. **Example** 1. Take **XLM-R** (Pre-trained on 100 langs). 2. Fine-tune on **English** Ner labels (Person, Org). 3. Run inference on **Arabic** text. 4. It successfully identifies Person/Org in Arabic. **Why It Works** - **Code-Switching**: The pre-training data often contains code-switching, linking languages. - **Shared Semantics**: The model learns that the *context* of a name looks similar across languages (structural alignment). - **Anchors**: Shared tokens (numbers, proper nouns, URLs) act as anchors to align the spaces. **Why It Matters** - **Benchmark**: The primary metric for evaluating multilingual models (XTREME benchmark). - **Magic**: One of the most surprising and useful emergent behaviors of deep learning. **Zero-Shot Cross-Lingual Transfer** is **instant polyglot skills** — learning a skill in English and immediately knowing how to do it in Arabic without practice.