For minoritized languages with scarce data, a typical approach is to pool texts from multiple dialects or varieties into one training corpus. A new arXiv paper challenges this practice, focusing on Nawatl (also spelled Nahuatl) and specifically the Masewalkopa variant. The authors argue that treating closely related varieties as a single corpus can fragment the model's learning, and they propose a single-variant specialization instead.

The paper, titled "Unfragmenting Nawatl NMT: Single-Variant Specialization for Masewalkopa," suggests that by narrowing the training data to one variant, the model can develop more consistent and accurate translations. The abstract notes that mixed corpora of closely related varieties are common, but implies this may be counterproductive for low-resource neural machine translation.

This is a single-source report, so there are no conflicting findings to compare. The full details of the method and evaluation are not available in the abstract, but the central claim is clear: for some low-resource languages, less variety in training data may lead to better results.