multilingual alignment

**Multilingual Alignment** is the **process or property of mapping representations from different languages into a shared vector space so that semantically similar words or sentences are close together regardless of language** — correcting the natural rotation or mismatch between independent language spaces. **Methods** - **Implicit**: Multilingual Masked Language Modeling (mBERT) creates implicit alignment. - **Explicit (Supervised)**: Use parallel corpora (translation pairs) and Minimize MSE($E_{eng}, E_{fr}$) — explicitly pulling translations together. - **TLM (Translation Language Modeling)**: Perform MLM on concatenated translation pairs, allowing the model to attend from English context to French target. **Why It Matters** - **Transfer Success**: Better alignment = better cross-lingual transfer. - **Retrieval**: Enables Cross-Lingual Information Retrieval (search French docs with English queries). - **Sentence Mining**: Used to find parallel sentences in noisy web crawls (like CommonCrawl) to build translation datasets. **Multilingual Alignment** is **synchronizing the maps** — ensuring the vector for "dog" in English lands on top of "perro" in Spanish in the high-dimensional embedding space.