CLD² Language Documentation Meets Natural Language Processing for Revitalising Endangered Languages

Roberto Zariquiey, Arturo Oncevay, Javier VeraACL 2022

TL;DR

This paper introduces CLD², a paradigm for integrating computational linguistics (NLP) with language documentation to revitalize endangered languages. It argues that current language documentation efforts produce valuable data that is underutilized by NLP, and proposes a framework for creating NLP-friendly annotated data as a standard deliverable of documentation projects.

What Problem Does This Paper Address?

The world faces a critical loss of linguistic diversity, with thousands of languages at risk of extinction. While documentary linguistics has made significant strides in creating comprehensive records of these languages, the resulting data often remains inaccessible or unusable for Natural Language Processing (NLP) applications. This disconnect means that the potential of NLP to aid in language revitalization—through tools like spell checkers, machine translation, or speech recognition—is largely untapped for endangered languages. The paper highlights a gap where language documentation produces data, but NLP struggles to leverage it, and conversely, NLP development for these languages is hindered by a lack of suitable data.

Methods

The paper proposes a new paradigm called Computational Language Documentation and Development (CLD²). This paradigm advocates for a two-way integration:

  1. NLP-friendly annotated data as deliverables: Future language documentation projects should aim to produce annotated data that is directly usable by NLP researchers. This includes:
  2. Monolingual and parallel corpora relevant to specific domains.
  3. Sentences annotated using universal frameworks like Universal Dependencies (UD) for syntax and Universal Morphology for morphology.
  4. Communication describing the characteristics of these annotations to aid NLP understanding.
  5. Exploitation of documentation databases by NLP: NLP researchers should actively utilize existing language documentation databases to develop NLP tools for endangered languages.

The authors also conducted an analysis of publication trends in NLP (ACL Anthology) and the coverage of endangered languages in documentation archives (ELAR) and multilingual datasets. This involved matching language names from these sources with the Glottolog database and classifying languages by their endangerment status.

Datasets

The paper discusses the use of several datasets and repositories:

  • ACL Anthology: A repository of NLP publications used to track research trends related to endangered languages.
  • Endangered Languages Archive (ELAR): A central database for language documentation projects, containing data for hundreds of languages.
  • Massively Multilingual (MM) Datasets: Including UniMorph, Universal Dependencies, and Tatoeba, used to assess the overlap with endangered language data.
  • Glottolog 4.4: An inventory of languages used for matching and geographical information.

The paper does not introduce a new benchmark dataset but rather analyzes the existing landscape to highlight the underutilization of data.

Key Insights and Ideas

  • The disconnect is a missed opportunity: Language documentation produces rich data, but its potential for NLP-driven revitalization is largely unrealized due to issues of accessibility, visibility, and readability for NLP researchers.
  • "Computerization" is key to revitalization: For endangered languages, "computerization"—the development of NLP tools—is a crucial step towards revitalization, not just a technological enhancement.
  • CLD² offers a practical bridge: The proposed CLD² paradigm provides a concrete framework for bridging the gap, suggesting actionable steps for both linguists and NLP researchers.
  • Readability is the biggest hurdle: While accessibility and visibility are challenges, the inherent structure and annotation schemes of language documentation data (even when machine-readable) are often not directly compatible with NLP tools, making "readability" the most difficult aspect to overcome.
  • Typological diversity benefits NLP: Access to data from typologically diverse endangered languages can help NLP researchers identify flaws in their methods and develop more robust, generalizable models.

Key Results

The analysis revealed a significant disparity:

  • While publications on endangered languages are growing, the number of NLP publications focusing on moribund or nearly extinct languages remains very low compared to those on non-endangered languages.
  • There is a very low overlap between languages documented in ELAR and those covered in the ACL Anthology or major multilingual NLP datasets. Specifically, out of 286 languages with geographical information in ELAR, only 22 (7.7%) appear in the ACL Anthology, and 12 (4.2%) in MM datasets.
  • Geographical areas like the Americas, Africa, and Southeast Asia are better covered by language documentation than by NLP resources, indicating potential for regional NLP initiatives to leverage these under-explored resources.

Limitations and Future Work

  • Readability of existing data: The paper acknowledges that much of the existing language documentation data, while machine-readable, is not easily processed by NLP tools due to inconsistent annotation schemas. This is a significant limitation that is hard to retroactively fix.
  • Cost-prohibitive massive development: The authors note that a complete overhaul of all documentation processes to include extensive technology development would be prohibitively expensive.
  • Future work suggestions:
    • Focus on IGT parsing and standardized schemas: Encourage research into parsing Interlinear Glossed Text (IGT) and establishing more universally readable IGT schemas.
  • Integrate NLP-friendly deliverables: Advocate for documentation projects to include NLP-friendly annotated data (corpora, UD/UM annotations) as standard outputs.
  • Develop NLP toolkits: Foster collaboration between NLPers and field linguists to create NLP toolkits for minority languages, leveraging existing databases and modern NLP techniques like transfer learning and multilingual pretraining.
  • Assess output quality: Emphasize the need to evaluate the quality of NLP tools developed for endangered languages to avoid misleading users.

Summary updated

BibTeX
@inproceedings{zariquiey2022cld,
  author = {Roberto Zariquiey and Arturo Oncevay and Javier Vera},
  title = {{CLD² Language Documentation Meets Natural Language Processing for Revitalising Endangered Languages}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2022},
}

Loading related content...

Figure