What Resources Matter for Interlinear Glossing? Using LLMs and RAG for the Low-Resource Mapudungun Language

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, medium

Summary

Research by Anaís Almendra, Arianna Bisazza, Claudio Gutierrez, and Felipe Hasler, presented at AmericasNLP 2026, explores using Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) for automatic interlinear glossing of Mapudungun, an endangered language spoken in Chile and Argentina. Their study, utilizing the Gemini 2.5 Pro model, investigated how different information configurations impact performance. They compared integrating a formal grammar, a dictionary, a small annotated corpus, and a combination of these resources. The evaluation revealed that while dictionary integration significantly degraded performance, grounding the model with a structured corpus maximized accuracy. Notably, a small dataset of 589 meaning units proved sufficient to substantially improve the morphological tagging task, demonstrating the viability of minimally annotated corpora for documenting morphologically complex languages.

Key takeaway

For NLP Engineers or computational linguists working on endangered or low-resource languages, you should prioritize developing even small, structured annotated corpora over relying on dictionaries for interlinear glossing tasks. Your efforts in creating a dataset of around 589 meaning units can significantly improve morphological tagging accuracy with LLMs like Gemini 2.5 Pro, making language documentation more efficient. Avoid integrating raw dictionaries, as this may degrade model performance.

Key insights

Minimally annotated corpora, not dictionaries, significantly enhance LLM-based interlinear glossing for morphologically complex languages.

Principles

• Structured corpora provide effective normative guidance.
• Dictionary integration can degrade LLM performance.
• Small datasets can yield significant linguistic improvements.

Method

The method involves configuring RAG with various linguistic resources (grammar, dictionary, corpus) to ground an LLM (Gemini 2.5 Pro) for automatic interlinear gloss generation.

In practice

• Use small annotated corpora for morphological tagging.
• Avoid direct dictionary integration in RAG for glossing.
• Apply RAG with structured data for low-resource languages.

Topics

• Interlinear Glossing
• Low-Resource Languages
• Mapudungun
• Large Language Models
• Retrieval-Augmented Generation
• Morphological Tagging

Best for: Research Scientist, AI Scientist, NLP Engineer, Domain Expert

Related on AIssential

• A Larger Annotated Corpus of Portuguese Coreference — LLMs can annotate Portuguese coreference, but a new LLM-based system did not outperform existing rule-based methods.
• AI-assisted cultural heritage dissemination: Comparing NMT and glossary-augmented LLM translation in rock art documents — Glossary-augmented LLMs significantly improve terminology accuracy in specialized translation without sacrificing overall quality.
• The Data Problem in Low-Resource Languages — Data availability, quality, and standardization are critical for NLP model performance across diverse languages.
• IndicDB -- Benchmarking Multilingual Text-to-SQL Capabilities in Indian Languages — IndicDB reveals a significant "Indic Gap" in LLM Text-to-SQL performance for Indian languages compared to English.
• When Similar Means Different: Evaluating LLMs on Arabic--Hebrew Cognates — Current LLMs struggle with cross-lingual semantic disambiguation, especially false friends and loanwords, due to over-reliance on surface forms.
• Vector RAG vs LLM-Compiled Wiki: A Preregistered Comparison on a Small Multi-Domain Research — Grounded research synthesis involves multiple capabilities, with no single architecture excelling across all metrics.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.