@InProceedings{fernandes:2026:latell,
  author    = {Fernandes, Rafael Macario},
  title     = {Cross-Lingual Transfer from Portuguese to Nheengatu: Evidence for Contact-Induced Convergence as a Computational Bridge},
  booktitle      = {Proceedings of the First International Conference on Language Technologies for Low-resource Languages (LaTeLL 2026)},
  month          = {September},
  year           = {2026},
  address        = {Fes, Morocco},
  publisher      = {Association for Computational Linguistics},
  pages     = {40--49},
  abstract  = {Typological distance predicts that Portuguese-to-Nheengatu transfer should fail: the languages are genetically unrelated and typologically divergent (URIEL+ genetic 1.0, syntactic 0.61). We argue that 400 years of contact-induced structural convergence may provide a computational bridge invisible to these measures, so that transfer should succeed at the levels where contact has operated most deeply. Across five experiments on a parallel corpus of about 5,000 sentence pairs, we find a progressive pattern: word-level and lexical-semantic methods fail, while fine-tuned XLM-RoBERTa reaches Precision@1 of 24.7\% and MRR of 0.371 on cross-lingual sentence retrieval—more than double a BM25 character-n-gram baseline and 1.7× that of zero-shot LaBSE, the strongest off-the-shelf retriever we test. Crucially, Portuguese beats Spanish—typologically near-identical and no less represented in pretraining, but without comparable contact history—on every metric and all five seeds, and the two sources share substantial lexical overlap, so lexical overlap alone is unlikely to explain the gap. These results suggest that contact history can distinguish transfer sources that typological distance treats as equivalent.},
  url       = {https://aclanthology.org/2026.latell-1.6}
}