Wdowiak customizes NLLB for sicilian translation
Wdowiak rebuilt a small neural machine translation system for Sicilian by extracting parallel text from the bilingual journal Arba Sicula and porting the original tokenization to a Perl‑free implementation. The new system was evaluated against a frozen test set from the journal.
Key points
- Wdowiak rebuilt Sicilian NMT using NLLB with low‑rank fine‑tuning
- Staged training with back‑translations improved Sicilian translation more than reverse training
- Multilingual adapter degraded performance on unseen language pairs
The study compared small Transformers trained from scratch with the pretrained NLLB model adapted by low‑rank fine‑tuning. The pretrained model performed better overall, and staged training—first forward‑scored back‑translations of web Sicilian, then curated literary text—improved translation into Sicilian more than reverse training.
On the test set, the staged model with a larger adapter and more back‑translations outperformed the reverse‑training recipe in English‑related directions but lagged between Sicilian and Italian. A multilingual adapter degraded performance on unseen language pairs, limiting its general use.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- Cohere releases Rerank 5 with passage-level context focus and calibrated scores · 1 src
- OpenAI says its AI agents weakened internal misalignment monitoring system · 31 src
- inclusionAI releases Ling-3.0-flash-Sante, a 124B healthcare model · 1 src
- Twelve AI models say humans prioritize belonging over truth, study finds · 1 src
- Recursive releases Flint, a 40B parameter MoE coding model with 2.5B active parameters per token · 6 src
Comments
via GitHub Discussions