Introduction
Language Identification (LangID) is the foundational step in modern natural language processing (NLP) pipelines. Without accurate LangID, downstream tasks like translation, search, and analysis cannot function. However, commercial tools frequently fail to recognize Native American languages, creating a significant digital divide for indigenous communities.
The Problem: Digital Exclusion and "Hallucination"
Major commercial LangID systems, such as Google Translate, often "hallucinate" when encountering Navajo (Diné Bizaad). Instead of identifying it as a Native American language, they mislabel it as unrelated languages like Icelandic, Lingala, or West African Wolof. This exclusion hinders language preservation and revitalization efforts. The authors argue that a decentralized, targeted approach is necessary to support underrepresented linguistic communities.
Methodology
The researchers implemented a Random Forest classifier with 100 trees. This model was chosen for its interpretability and robustness on smaller, specialized datasets compared to resource-heavy deep learning models.
- Feature Extraction: Input text was vectorized into 5,000-dimensional features to capture a wide range of linguistic attributes.
- Classification Task: The model was trained to distinguish Navajo from the 20 languages most frequently confused by commercial tools.
- Datasets: The study utilized 10,000 Navajo sentences and a "confusion" dataset of 20 unrelated languages (e.g., Icelandic, Czech, Yoruba), all sourced from the Leipzig corpora.
Key Insights: The Family Tree Effect
A significant finding was the model's ability to generalize to related languages. Although trained only to recognize Navajo, the model frequently identified Apache sentences as Navajo. This "zero-shot" behavior aligns perfectly with the Athabaskan language family structure.
The model demonstrated high sensitivity to the "Western Apachean" subgroup—which includes Navajo, Western Apache, and Mescalero—identifying them as Navajo 96-100% of the time. In contrast, it was less likely to mistake languages from the "Eastern" subgroup (Jicarilla or Lipan Apache) for Navajo, mirroring linguistic reality.
Key Results
The Random Forest classifier achieved an overall accuracy of 97% across the 21-language test set. For Navajo specifically, the model reached a precision of 1.00 and a recall of 0.99, a massive improvement over the 0% accuracy of commercial tools on the same data. Furthermore, the generalization test showed that 100% of Mescalero Apache and 96% of Western Apache sentences were identified as Navajo, proving the model's utility as a broad Athabaskan detector.
Conclusion
This work demonstrates that high-accuracy detection for endangered languages is possible even with limited resources. By moving away from centralized "one-size-fits-all" models toward localized, culturally inclusive NLP, we can better support linguistic revitalization and bridge the digital divide for indigenous communities.
Summary updated