BELT: Building Endangered Language Technology

Michael Ginn, David Saavedra-Beltrán, Camilo Robayo, Alexis PalmerACL 2024

TL;DR

This paper introduces BELT (Building Endangered Language Technology), a free, open-source web platform designed to empower speakers of endangered languages to build their own language technologies. By offering interactive Python lessons and project-based tutorials, BELT aims to democratize NLP development and foster data sovereignty within language communities, moving away from colonialist approaches.

What Problem Does This Paper Address?

The development of language technologies (LTs) for endangered languages, while often seen as a goal for language revitalization, is fraught with social and ethical challenges. Historically, such efforts have often been "colonialist," characterized by the extraction of language data without regard for speaker privacy, a lack of community sovereignty over their data, a focus on tools that benefit external experts rather than speakers, and reliance on outside experts. Furthermore, NLP research often underrepresents these languages, and existing resources are not tailored to the specific needs and contexts of endangered language communities, who may lack extensive technical backgrounds.

Methods

BELT is built upon the JupyterHub platform, providing users with individual Jupyter environments. The core of BELT consists of:

  • Interactive Python Lessons: These lessons cover fundamental programming concepts and are specifically framed within the context of language technology. They are designed to be approachable for beginners.
  • Project-Based Tutorials: These longer tutorials guide users through building practical language technologies like spellcheckers, predictive text models, and word games. These projects are language-agnostic, requiring only text data as input.
  • Gradio Framework Integration: For deploying the developed technologies, BELT utilizes the Gradio framework. This allows users to create simple, interactive user interfaces for their applications, which can be temporarily accessed via a live URL or permanently hosted on platforms like Hugging Face Spaces.
  • Localization: BELT is designed for global accessibility, with preliminary localization into Spanish and Portuguese. A JavaScript helper is used to dynamically switch language content.

Datasets

The paper does not explicitly list specific datasets used for training or evaluation. However, it mentions that the project tutorials are designed to be language-agnostic and encourage learners to use their own language data. Examples of languages used in project tutorials include English, Spanish, Latin, and Uspanteko.

Key Insights and Ideas

  • Democratizing NLP for Endangered Languages: The central idea is to shift the power of LT development from external experts to the language communities themselves, promoting self-determination and data sovereignty.
  • Focus on Practical, Low-Resource Technologies: BELT prioritizes building tools that are computationally inexpensive, require minimal data, and do not necessitate extensive annotation, making them more feasible for endangered language contexts.
  • Interactivity is Key: The use of Jupyter notebooks and interactive exercises is crucial for engaging learners and facilitating a hands-on learning experience.
  • Bridging the Gap Between Linguistics and Technology: The platform aims to equip linguists and community members with the technical skills needed to create practical tools for their languages.
  • Importance of Localization and Accessibility: Making the platform available in multiple languages is essential for reaching a wider audience and ensuring inclusivity.

Key Results

The paper reports on two live workshops conducted in South America:

  • Bogotá, Colombia (June 2023): Attended by fifteen participants, mostly linguistics/anthropology undergraduates. The workshop successfully introduced participants to BELT's core concepts and projects. Participants were able to create applications for languages like Karijona and Umbra, including word games. However, time constraints limited exploration of more advanced topics like Regular Expressions and Machine Learning.
  • Santiago, Chile (January 2024): This workshop had over 40 participants from diverse backgrounds (linguists, Indigenous language speakers, computer scientists). It focused on demonstrating the attainability of producing simple language technologies. Participants were able to build word games for four Indigenous languages (Mapudungan, Aymara, Quechua, and Ckunsa) which were then hosted on Hugging Face Spaces.

Qualitatively, participants found the materials positively received and expressed that seeing how programming languages work opened new possibilities for their languages.

Limitations and Future Work

  • Limitations:
    • The current implementation of the localization framework may not scale well to a very large number of components.
  • While the developed tools are suitable foundations, they may not be immediately suitable for real-world use depending on the input data.
  • Limited internet access in some communities can hinder regular use of online applications.
  • Future Work:
    • Mobile Version: Developing a mobile version of BELT to enable offline use of applications, especially for younger community members.
  • Broader Localization: Translating course materials into a wider range of languages of wider communication.
  • Additional Skills and Projects: Expanding the platform with more content.
  • Self-Hosting: Exploring the possibility of allowing tribal governments, immersion schools, or summer schools to self-host BELT access.
  • Community Contributions: Welcoming contributions from the broader NLP community for new projects and content.

Summary updated

BibTeX
@inproceedings{ginn2024belt,
  author = {Michael Ginn and David Saavedra-Beltrán and Camilo Robayo and Alexis Palmer},
  title = {{BELT: Building Endangered Language Technology}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
}
Figure