$\mu$\textsc{gnaw}$\oplus$(0,1,2): three context-free micro-grammars oriented for machine learning of Nawatl language

ACL ARR 2025 July Submission300 Authors

Abstract:

We present in this article three context-free micro-grammars (CFMG) for the Nawatl language.
Nawatl, an Amerindian language, is a $\pi$-language, or a low-resource language in terms of digital data, and the corpora available for Large Language Models (LLMs) are virtually nonexistent.
The objective is to generate a significant number of Nawatl sentences in order to expand the available corpus for training static embeddings or LLMs.
Using the best micro-grammar, we significantly expanded the Nawatl corpus $\pi$-yalli, then we used this enriched corpus to train FastText and applied it to a sentence-level semantic task.
The results show an improvement compared to results obtained using only the original corpus without artificial expansion, which is encouraging.

Paper Type: Long

Research Area: Efficient/Low-Resource Methods for NLP

Research Area Keywords: Nawatl language, Data augmentation, LLM Efficiency, NLP in resource-constrained settings; Statics Embeddings

Contribution Types: Approaches to low-resource settings, Data resources

Languages Studied: Nahuatl (Amerindian Language)

Submission Number: 300