romanization

**Romanization** is a **specific type of transliteration where text is converted into the Roman (Latin) script** — often used in NLP to standardize inputs from diverse scripts (Cyrillic, Arabic, Devanagari, CJK) into a common character set to facilitate transfer or vocabulary sharing. **Usage in NLP** - **Vocabulary Reduction**: Instead of a vocabulary of 200,000 (learning Chinese chars, Hindi chars, etc.), Romanization allows a smaller shared vocabulary. - **Alignment**: It's easier for a model to see that "Bank" (En) and "Bank" (De) are similar if they share the script. Not so for "Bank" vs "банк". - **u-PMLM**: Universal Phonetic/Romanized models convert all training data to IPA or Latin script to focus on phonetic/semantic similarities. **Why It Matters** - **Low-Resource**: Helps languages with rare scripts benefit from high-resource Latin-script transfer. - **Lossy**: Romanization typically loses information (tonal marks, distinct spellings mapping to same phoneme) — inevitable trade-off. **Romanization** is **standardizing to Latin** — converting all world scripts into A-Z characters to maximize overlap and sharing in multilingual models.