Mining Grammar Books with OCR and VLMs to Build Translation Data for 4000+ Low-Resource Languages

A new paper by Chutong Meng and Barron Associates extracts parallel corpora from grammar books and archival texts using OCR and VLMs to improve translation for extremely low-resource languages. With just 512 mined sentences, Qwen's average chrF++ on 10 unsupported languages jumped from 9.36 to 30.76.

2026-11-06 ~ 2026-11-06 · 2 related posts

• OCR+VLM pipeline mines grammar books to boost translation for 4,000+ low-resource languages — anas_ant · 2026-11-06
• 512 mined sentence pairs lift Qwen's low-resource MT by +21.4 chrF++ on average — anas_ant · 2026-11-06