Chittagonian · ctg · Bangla dialect
বাংলা → চাটগাঁইয়া
Converts standard Bangla to Chittagonian using 656 phonological rules mined automatically from parallel data — ক→গ, প→ফ, খ→হ, হ→অই — plus a 3,903-entry lexicon. No hand-written linguistics, and nothing leaves your browser.
| Bangla | Chittagonian | via |
|---|
What this is, and is not. It is the rule system that generates training data for ChatgaiyyaLM, not a trained model. On held-out Bangla→Chittagonian it scores 57.94 chrF — higher than the best adapted neural model in the same project (52.63), at no GPU cost. That is reported rather than hidden: a rule system beating a fine-tuned LM says something real about how much signal a small, clean lexicon carries when data is this scarce.
Rules alone reach 35.15 chrF, and that is the honest number for vocabulary the miner never saw — the lexicon short-circuits words it already knows.
Words often pass through unchanged, and that is usually correct: many Bangla words genuinely are identical in Chittagonian. The rule ^[প→ফ]া has precision 0.60 precisely because প stays প about 40% of the time in that position.