On October 24, 2026, a research study on the reconstruction of a neural machine translation (NMT) system for Sicilian was published on arXiv (paper ID 2610.30374v1). The study aims to improve the translation quality of current multilingual large models—currently trained on noisy network text—by redeveloping Wdowiak’s earlier small-scale NMT system customized for this language.
The team, leveraging the bilingual academic journal “Arba Sicula” in PDF format, rebuilt the parallel corpus extraction pipeline using well-maintained software: they employed a language-agnostic sentence encoder for sentence alignment, with thresholds calibrated against manually aligned results and typical automatic alignment errors filtered out; meanwhile, the original system’s Perl-dependent tokenization and inflection-sensitive subword segmentation approach was migrated to a Perl-free implementation. All experimental results were evaluated on a frozen test set derived from this journal.
The study compared a small Transformer model trained from scratch with a pre-trained NLLB model fine-tuned using low-rank adaptation (LoRA), in bidirectional English↔Sicilian and Italian↔Sicilian tasks, with Italian serving as a closely related bridge language. The authors also transferred the original system’s “reverse training” strategy—first translating Sicilian back into Italian for scoring, then incorporating the literary corpus of “Arba Sicula”—to the NLLB model. The results show that the pre-trained NLLB model outperformed the from-scratch model across the board, and that the staged training approach yielded particularly pronounced improvements when translating into Sicilian.
On the original system’s test set and within the original tokenization space, the staged model—equipped with a larger adapter and more back-translation data—outperformed the original reverse-training approach in the English↔Sicilian direction, but lagged slightly in the Sicilian↔Italian direction. Furthermore, multilingual adapters degrade performance on language directions not involved in training, thereby limiting their applicability as general-purpose multilingual models.