Climbing the Ladder: Automating Data Pipelines and Reproducing Baselines in Low-Resource Settings

A new arXiv paper (2610.30374v1) reports that a pretrained NLLB model adapted with low-rank fine-tuning substantially outperforms small Transformers trained from scratch for Sicilian machine translation, with staged training on forward-scored back-translations of web Sicilian followed by curated literary text improving translation into Sicilian more than translation from it. The authors rebuilt the data pipeline of Wdowiak's original Sicilian NMT system on maintained software, re-extracting parallel text from PDFs of the bilingual journal Arba Sicula, aligning sentences with a language-agnostic sentence encoder tuned against an existing hand alignment, and porting the tokenization and desinence-biased subword segmentation to a Perl-free implementation, all measured on a frozen test set drawn from the journal. On the original system's test set and tokenized space, the staged model with a larger adapter and more back-translations beats the reverse-training recipe in directions involving English but falls below it between Sicilian and Italian, and a multilingual adapter degrades directions it was not trained on, limiting its use as a general multilingual model.

arXiv:2610.30374v1 Announce Type: new Abstract: The massively multilingual translation systems that now include Sicilian learned it from noisy, non-standardised web text. To obtain better translation quality, we redevelop the small neural machine translation NMT system that Wdowiak customized for Sicilian. We rebuild on maintained software the extraction of the parallel text from the PDFs of the bilingual journal Arba Sicula, whose sentences are aligned by a language-agnostic sentence encoder with thresholds tuned against an existing hand alignment and then filtered for the typical errors of automatic alignment. Furthermore, the tokenization and the desinence-biased subword segmentation of the original system are ported to a Perl-free implementation, and all the results are measured on a frozen test set drawn from the journal. We compare small Transformers trained from scratch with the pretrained NLLB model adapted by low-rank fine-tuning, in both directions and with Italian as a closely related bridge language, and we adapt to the pretrained model the reverse-training recipe of the original system, i.e., a stage on forward-scored back-translations of the web Sicilian followed by a stage on the curated literary text. The pretrained model is by far the better of the two, and the staged training improves the translation into Sicilian more than the translation from it. On the test set of the original system, and in the tokenized space in which it was scored, the staged model with a larger adapter and more back-translations is above the reverse-training recipe in the directions that involve English and below it between Sicilian and Italian. Yet, a multilingual adapter degrades the directions on which it was not trained, which limits its use as a general multilingual model.