Bringing Mapudungun into the Modern MT Ecosystem: Morphology-Aware Tokenization for NLLB-200 Fine-Tuning

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

Researchers fine-tuned Meta's NLLB-200 for Mapudungun–Spanish translation, demonstrating that morphology-aware tokenization can significantly reduce model size requirements. Their novel Morfessor-VC method, which constrains Morfessor morpheme segmentation to NLLB's pretrained vocabulary, enabled a 600M parameter model to achieve 43.2 chrF++ for Mapudungun→Spanish translation. This performance is competitive with a 3.3B parameter model using Standard BPE (42.9 chrF++) while using five times fewer parameters. All fine-tuned conditions surpassed frontier LLMs by over 27 chrF++. This work establishes new baselines for Mapudungun, an indigenous polysynthetic language spoken by 200,000+ Mapuche people, which is absent from NLLB-200 and commercial MT providers. It addresses tokenization challenges posed by Mapudungun's agglutinative morphology, providing a practical foundation for community language tools.

Key takeaway

For NLP Engineers developing machine translation for low-resource or morphologically complex languages, you should prioritize morphology-aware tokenization. This approach, exemplified by Morfessor-VC, allows you to achieve competitive performance with significantly smaller models, like a 600M parameter NLLB-200, compared to much larger 3.3B models. Consider integrating custom tokenization strategies to optimize resource usage and expand MT support for underserved languages.

Key insights

Morphology-aware tokenization significantly reduces model size for low-resource, polysynthetic language machine translation while maintaining high performance.

Principles

• Morphology-aware tokenization improves MT efficiency.
• Pretrained vocabulary constraints enhance segmentation.
• Low-resource languages benefit from specialized tokenization.

Method

Fine-tuning NLLB-200 on Mapudungun–Spanish translation using eight tokenization strategies, including Morfessor-VC, which constrains morpheme segmentation to NLLB's pretrained vocabulary.

In practice

• Develop MT for polysynthetic languages.
• Create tools for language revitalization.
• Build pedagogical resources for Mapudungun.

Topics

• Machine Translation
• Mapudungun Language
• NLLB-200
• Morphology-Aware Tokenization
• Low-Resource Languages
• Language Revitalization

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Modelling the Morphology of Verbal Paradigms: A Case Study in the Tokenization of Turkish and Hebrew — Tokenization strategy critically impacts Transformer models' ability to represent complex morphology, especially in non-concatenative languages.
• Surpassing Scale by Efficiency: A Compact 135M Parameter Foundational LLM Natively Adapted for the Bangla Language — A compact 135M parameter LLM for Bangla achieves 1B-tier performance by efficiently merging tokens, overcoming low-resource script challenges.
• Multilingual Coreference Resolution via Cycle-Consistent Machine Translation — Cycle-consistent machine translation effectively generates training data for low-resource multilingual coreference resolution, improving performance where corpora are scarce.
• Equity with Efficiency: An Empirical Study of Tokenizers for Multilingual Large Language Models — Cross-lingual fairness and tokenization efficiency are not fundamentally at odds in multilingual LLMs.
• IHUBERT: Vector-Based Semantic Deduplication and Domain-Balanced Pretraining for Persian Resources — Semantic deduplication and domain-balanced pretraining significantly enhance Persian PLM performance across diverse NLU tasks.
• Disentangling meaning from language in LLM-based machine translation — LLM machine translation functions are disentangled into distinct, specialized attention head sets for language and meaning.

Open in AIssential →

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