Balancing Social Impact, Opportunities, and Ethical Constraints of Using AI in the Documentation and Vitalization of Indigenous Languages

Claudio S. Pinhanez, Paulo R. Cavalin, Marisa Vasconcelos, Julio NogimaIJCAI 2023 · DOI

TL;DR

This paper explores the critical intersection of Artificial Intelligence (AI) and the preservation of Indigenous languages. It argues that using AI for language documentation and revitalization offers a unique opportunity to address both societal needs and AI's own technical and ethical shortcomings, advocating for a balanced approach that prioritizes community involvement and ethical considerations.

What Problem Does This Paper Address?

The paper addresses the urgent challenge of Indigenous language endangerment, with a significant portion of the world's languages at risk of extinction. It highlights the conflict between traditional Indigenous cultures and the rapid advancement of technology, particularly AI and Natural Language Processing (NLP). The authors identify a critical gap in how AI is developed and deployed, often relying on practices like data extractivism and colonial thinking, which are antithetical to the principles of Indigenous communities. They propose that the specific needs and constraints of working with Indigenous languages can serve as a catalyst for developing more ethical, robust, and socially responsible AI.

Methods

The paper doesn't present a single novel algorithm but rather a framework and discussion around the ethical and practical application of existing AI/NLP techniques.

  • Survey of Existing Work: The authors review previous research on using AI and NLP for language documentation and vitalization, identifying key limitations of current technologies.
  • Ethical Frameworks: They discuss established ethical guidelines for engaging with Indigenous communities, emphasizing principles like native-centrism, respect, and community ownership.
  • Case Study (Prototype LLM): A significant part of the paper details the development of a prototype Large Language Model (LLM) for Brazilian Indigenous languages (BILs). This involved training a modified BERT model (mBERT) on translated Bible texts from 51 BILs and three high-resource languages. The goal was to test the feasibility of using LLMs for low-resource languages.
  • Discussion of Trade-offs: The paper emphasizes the inherent trade-offs between achieving social impact, leveraging technical opportunities, and adhering to ethical constraints, particularly in the context of limited data and resources.

Datasets

  • Bible Translations: The primary dataset used for training the prototype LLM for Brazilian Indigenous languages consisted of publicly available translations of the Bible into 51 BILs.
  • High-Resource Languages: Translations from Portuguese, Spanish, and English were also included to facilitate comparison and understanding of linguistic similarities and differences.
  • Existing Corpora: The paper references the use of various corpora for language documentation and NLP tasks, though specific details for Indigenous languages are often noted as scarce.

Key Insights and Ideas

  • AI for Good as a Catalyst for AI Improvement: The domain of Indigenous language documentation and vitalization is presented not just as a beneficiary of AI, but as a crucial testing ground that can force AI to confront its own limitations, such as data addiction, lack of explainability, and reliance on colonial mindsets.
  • The "Monster" LLM as a Necessary Evil: The authors acknowledge that their prototype LLM development likely violated ethical guidelines by not involving communities upfront. However, they argue this was a pragmatic step to demonstrate the potential of LLMs for low-resource languages, with strict containment measures to prevent misuse. This highlights the difficult "balancing act" required.
  • "Nothing for Us Without Us" as a Guiding Principle: This Indigenous motto is central to the paper's argument, underscoring the necessity of community-led initiatives, co-design, and genuine partnership in any AI development for these communities.
  • Decolonizing AI is Essential: The paper strongly advocates for moving beyond colonialist practices in AI development, emphasizing Indigenous leadership, diverse co-design, and a focus on community culture and needs.
  • Indigenous Languages as a Driver for Technical Innovation: The unique characteristics of Indigenous languages (low data, distinct linguistic structures, non-traditional contexts) can push AI to develop solutions for challenges like zero-shot learning, better context handling, and more culturally-aware common-sense reasoning.

Key Results

  • Prototype LLM Performance: The trained mBERT model for Brazilian Indigenous languages achieved approximately 95% accuracy on a language identification task, validating the hypothesis that LLMs can be viable foundational tools for low-resource languages.
  • Ethical Challenges Identified: The development process highlighted significant ethical hurdles, including the lack of community involvement in initial design, data sourcing issues (using religious texts), and unclear community needs for the resulting tools.
  • Demonstration of Trade-offs: The paper illustrates that achieving social impact often necessitates difficult compromises between technical feasibility and ethical purity, especially in the early stages of research.

Limitations and Future Work

  • Ethical Violations in Prototype Development: The authors explicitly acknowledge that their initial LLM prototype development likely violated ethical guidelines by not involving Indigenous communities from the outset.
  • Data Scarcity: The fundamental challenge of limited data for most Indigenous languages remains a significant hurdle for developing robust AI models.
  • Generalizability of LLMs: While the prototype showed promise, the paper notes it's unclear if multi-lingual models trained on web data can perform well with the distinct linguistic roots of Indigenous languages.
  • Future Work: The authors are actively pursuing collaborations with Indigenous communities for co-design and development of writing-support tools. They also aim to develop a sustainable work and governance model for LLM development for BILs in partnership with relevant organizations and Indigenous leadership. They emphasize the need for continued discussion on admissible trade-offs and the elaboration of design and development guidelines for AI in this domain.

Summary updated

BibTeX
@inproceedings{pinhanez2023balancing,
  author = {Claudio S. Pinhanez and Paulo R. Cavalin and Marisa Vasconcelos and Julio Nogima},
  title = {{Balancing Social Impact, Opportunities, and Ethical Constraints of Using AI in the Documentation and Vitalization of Indigenous Languages}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.24963/ijcai.2023/685},
}
Figure