TLDR: Researchers have developed a context-free grammar for Nawatl, an indigenous Mexican language with limited digital resources. This grammar generates artificial, grammatically correct sentences, which are then filtered for semantic plausibility and combined with an existing Nawatl corpus. This augmented dataset significantly improves the performance of static language models like FastText on semantic similarity tasks, demonstrating a promising method for revitalizing and developing computational tools for low-resource languages.
In the realm of language technology, the development of robust tools for languages with limited digital resources, often termed ‘π-languages,’ presents a significant challenge. A recent research paper introduces a novel approach to address this issue for Nawatl, a prominent indigenous language of Mexico, by employing a context-free grammar (CFG) to augment its digital corpora.
Nawatl, also known as Nahuatl, is spoken by approximately 1.65 million people and is characterized by substantial dialectal diversity. Despite its cultural importance, digital resources for Nawatl are scarce, making it difficult to train modern machine learning models for language processing. This scarcity motivated researchers to explore methods for artificially expanding existing Nawatl text collections.
A Grammar-Based Approach to Corpus Expansion
The core of this research lies in the creation of a context-free micro-grammar, dubbed µGNAW⊕0, specifically tailored for Nawatl. Unlike grammars designed for Indo-European languages, this micro-grammar acknowledges Nawatl’s agglutinative and polysynthetic nature, where words are formed by joining multiple morphemes to a root. While simplified, focusing on singular forms and present tense, the grammar’s primary goal is not to fully model the language’s complexity but to generate a large volume of syntactically valid sentences.
The generative capacity of µGNAW⊕0 allowed for the creation of hundreds of thousands of artificial sentences. However, simply generating grammatically correct sentences isn’t enough; they also need to be semantically plausible. To achieve this, the researchers implemented a series of semantic filters. These filters, which include an animate/inanimate noun-verb association and a check for repeated nouns, ensure that the generated sentences make sense, avoiding nonsensical phrases like “The big corn-cob eats a lot of rabbit.”
Augmenting the π-YALLI Corpus
The filtered artificial sentences were then merged with an existing authentic Nawatl corpus called π-YALLI. This combined dataset, named π-YALL-IA⊕0, underwent a standardization process to unify spellings across different Nawatl variants. The original π-YALLI corpus, while valuable, contains a relatively small number of tokens (around 6.1 million) and sentences (around 333,000), which is often insufficient for training advanced language models effectively.
The augmentation proved substantial, with the artificial text contributing approximately 70% of the tokens and an astounding 193% of the sentences compared to the authentic π-YALLI v1.10 corpus. This significant increase in data volume is crucial for improving the learning capabilities of language models.
Evaluating Semantic Similarity
To assess the impact of this corpus augmentation, the researchers trained FastText static embeddings using the expanded π-YALL-IA⊕0 corpus. These embeddings were then evaluated on a sentence-level semantic similarity task, where models had to rank candidate sentences based on their semantic closeness to a reference sentence. The performance was measured using Kendall’s τ, a metric for comparing rankings.
The results were encouraging. The FastText model trained on the filtered π-YALL-IA 1.10⊕0 corpus achieved a Kendall’s τ of 0.527. This placed it third in a comparative ranking against several large language models (LLMs) accessed via APIs, including Gemini 2.5 and Claude 3.7, and interactive models like Copilot and Grok 3. This demonstrates that even a simplified context-free grammar can significantly enhance the performance of static language models for low-resource languages.
Also Read:
- Advancing Kurdish Language Understanding with a New Semantic Textual Similarity Dataset
- Evaluating RAG Systems with Knowledge Graphs: A New Framework for Deeper Semantic Analysis
Future Directions and Impact
While the current micro-grammar shows promising results, the researchers acknowledge the inherent complexities of Nawatl as a polysynthetic language. Future work aims to develop a more sophisticated CFG that more accurately reflects Nawatl grammar, moving beyond Indo-European inspirations. This includes expanding the vocabulary, incorporating more grammatical persons, verb tenses, plurals, and rhetorical connectors, and refining the semantic filters.
This research highlights a viable path for the computational development and revitalization of low-resource languages, particularly those within the Uto-Aztecan family. By generating synthetic data, it addresses the critical lack of digital resources, paving the way for more advanced NLP applications like sentiment analysis, text summarization, and named entity recognition for Nawatl. For more details, you can read the full paper here.
