Dialectal machine translation remains a weak spot for current systems, largely because training data is scarce and dialects vary far more than standardized language. Standard benchmarks tend to assume well-edited, uniform text, so they fail to capture the linguistic variation that makes dialects like Silesian difficult to translate.
The paper, posted on arXiv, introduces a new Polish-Silesian benchmark and a translation system built for this specific pair. According to the abstract, the system outperforms both open-source and commercial translation models on this task, despite the common assumption that larger scale is the main driver of quality.
Because this is a single preprint, the details of the benchmark construction and system architecture are not yet fully available from the abstract alone. Still, the result points to the value of precision—targeted data and evaluation—over sheer model size for dialectal translation.