rmacario/nhengatu-experiments

Cross-Lingual Transfer: Portuguese → Nheengatu (experiments)

Models, embeddings, and results from a study of cross-lingual transfer between Portuguese (pt) and Nheengatu (yrl), a Tupian language of the Brazilian Amazon. Four methods are tested at increasing linguistic depth (Word2Vec, fastText, Procrustes alignment, and fine-tuned XLM-RoBERTa), plus a Portuguese-vs-Spanish source-language control.

Repository contents

Folder	Contents
01_word2vec	Word2Vec embeddings and analysis
02_fasttext	fastText (subword) embeddings
03_crosslingual	XLM-R fine-tuning, Procrustes alignment, and the PT-vs-ES control (scripts + results)
05_visualization	Figures (PCA, semantic fields)
06_typological_analysis	URIEL/typological distance analysis
results	Aggregated result files

Corpus

4,997 cleaned Portuguese–Nheengatu sentence pairs (5,028 before filtering) from four sources:

Source	Register	Pairs
Brazilian Constitution translated into Nheengatu	legal	2,020
Navarro, Curso de Língua Geral	pedagogical	2,139
Tycho Brahe historical fragments	historical	597
Casasnovas, Noções de língua geral	grammatical	241
Total (cleaned)		4,997

Navarro's pedagogical grammar and Casasnovas's grammatical notes were both obtained via the CompLin/nheengatu repository (Alencar, 2023); the Tycho Brahe fragments come from the Tycho Brahe corpus (Veronesi, 2026).

Results

Cross-lingual sentence retrieval, fine-tuned xlm-roberta-base, on a 1,000-pair held-out test set:

Metric	Score
Precision@1	0.247
Precision@5	0.504
Precision@10	0.621
MRR	0.371

By domain (P@1): Tycho 0.387 · Constitution 0.252 · Navarro 0.218 · Casasnovas 0.057

Source-language control

To test whether the advantage is contact-specific rather than generic multilingual behaviour, Portuguese is compared against Spanish — typologically near-identical but without comparable contact history — under identical fine-tuning across 5 seeds:

Metric	PT (contact)	ES (control)	Gap
P@1	0.281 ± 0.012	0.232 ± 0.005	+0.050
P@5	0.545 ± 0.034	0.453 ± 0.009	+0.091
P@10	0.673 ± 0.030	0.568 ± 0.006	+0.104
MRR	0.406 ± 0.018	0.340 ± 0.006	+0.066

Portuguese wins on every metric across all 5 seeds. Because the Spanish set is machine-translated from the Portuguese source, the gap is a conservative estimate of the contact-specific advantage.

Related

• Fine-tuned model: https://huggingface.co/rmacario/nhengatu-xlmr
• Code: https://github.com/rmaacario/nheengatu-crosslingual

Citation

@inproceedings{fernandes2026nheengatu,
  author    = {Fernandes, Rafael Macário},
  title     = {Cross-Lingual Transfer from Portuguese to Nheengatu: Contact-Induced Convergence as a Computational Bridge},
  booktitle = {Proceedings of LaTeLL 2026},
  year      = {2026}
}

License

CC BY 4.0 — Rafael Macário Fernandes.