Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Diacritic Restoration for Low-Resource Indigenous Languages: Case Study with Bribri and Cook Islands Māori

Rolando Coto-Solano, Daisy Li, Manoela Teleginski Ferraz, Olivia Sasse, Cha Krupka, Sharid Loáiciga, Sally Akevai Tenamu Nicholas

TL;DR

The paper tackles diacritic restoration in two extremely under-resourced languages, Bribri and Cook Islands Māori, by comparing statistical, fine-tuned LLM, and zero-shot approaches. It finds that fine-tuned, character-level models like ByT5 achieve the best restoration performance, with reliable results emerging at around 10,000 words of training data; zero-shot approaches perform poorly. The study analyzes diacritic-specific errors, data scale effects, and cross-language transfer against higher-resource languages, highlighting the benefits and limits of multilingual models. It also addresses diacritic correction and data sovereignty concerns, emphasizing community involvement and responsible use of language technologies in indigenous contexts.

Abstract

We present experiments on diacritic restoration, a form of text normalization essential for natural language processing (NLP) tasks. Our study focuses on two extremely under-resourced languages: Bribri, a Chibchan language spoken in Costa Rica, and Cook Islands Māori, a Polynesian language spoken in the Cook Islands. Specifically, this paper: (i) compares algorithms for diacritics restoration in under-resourced languages, including tonal diacritics, (ii) examines the amount of data required to achieve target performance levels, (iii) contrasts results across varying resource conditions, and (iv) explores the related task of diacritic correction. We find that fine-tuned, character-level LLMs perform best, likely due to their ability to decompose complex characters into their UTF-8 byte representations. In contrast, massively multilingual models perform less effectively given our data constraints. Across all models, reliable performance begins to emerge with data budgets of around 10,000 words. Zero-shot approaches perform poorly in all cases. This study responds both to requests from the language communities and to broader NLP research questions concerning model performance and generalization in under-resourced contexts.

Diacritic Restoration for Low-Resource Indigenous Languages: Case Study with Bribri and Cook Islands Māori

TL;DR

The paper tackles diacritic restoration in two extremely under-resourced languages, Bribri and Cook Islands Māori, by comparing statistical, fine-tuned LLM, and zero-shot approaches. It finds that fine-tuned, character-level models like ByT5 achieve the best restoration performance, with reliable results emerging at around 10,000 words of training data; zero-shot approaches perform poorly. The study analyzes diacritic-specific errors, data scale effects, and cross-language transfer against higher-resource languages, highlighting the benefits and limits of multilingual models. It also addresses diacritic correction and data sovereignty concerns, emphasizing community involvement and responsible use of language technologies in indigenous contexts.

Abstract

We present experiments on diacritic restoration, a form of text normalization essential for natural language processing (NLP) tasks. Our study focuses on two extremely under-resourced languages: Bribri, a Chibchan language spoken in Costa Rica, and Cook Islands Māori, a Polynesian language spoken in the Cook Islands. Specifically, this paper: (i) compares algorithms for diacritics restoration in under-resourced languages, including tonal diacritics, (ii) examines the amount of data required to achieve target performance levels, (iii) contrasts results across varying resource conditions, and (iv) explores the related task of diacritic correction. We find that fine-tuned, character-level LLMs perform best, likely due to their ability to decompose complex characters into their UTF-8 byte representations. In contrast, massively multilingual models perform less effectively given our data constraints. Across all models, reliable performance begins to emerge with data budgets of around 10,000 words. Zero-shot approaches perform poorly in all cases. This study responds both to requests from the language communities and to broader NLP research questions concerning model performance and generalization in under-resourced contexts.
Paper Structure (25 sections, 3 figures, 11 tables)

This paper contains 25 sections, 3 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Diacritic Restoration
  4. Orthographic Correction and Diacritics
  5. Cook Islands Māori: Long vowels and glottal stops
  6. Bribri: Tones, nasals and glottal stops
  7. Evaluation of Different Approaches for Diacritic Restoration
  8. Data Inputs
  9. Models
  10. Benchmarking Model Effectiveness
  11. Results
  12. Confusion patterns in complete restoration
  13. Amounts of Data and Performance
  14. Results
  15. Comparison with High-Resource Languages
  16. ...and 10 more sections

Figures (3)

  • Figure 1: Example of typewritten text in CIM with long vowel macrons and glottal stop saltillos absent Ponga:1975. In the Nicholas2017manual orthography, the first words would be Tāna 'anga'anga 'her work'.
  • Figure 2: Example of spelling variation in Bribri ietsay-narraciones. Using Constenla's cursobribri orthography, the first two lines would be: Tkabë̀köl ditsèwö tskìna: Tkabë̀köl ditsèwö tskìna e' aláköl amì akë̀. "The Snake People are Created: The Snake People who were created (came from) a woman, (her) mother (and her) brother". For example, the word ditsèwö 'clan, people' appears as desetwö and detséwö.
  • Figure 3: WER of ByT5 outputs as a function of word presence in the training set.