Proceedings of the Workshop on Natural Language Processing for Indigenous Languages of the Americas (AmericasNLP)

Manuel Mager, Abteen Ebrahimi, Arturo Oncevay, Enora Rice, Shruti Rijhwani, Alexis Palmer, Katharina KannACL 2023

TL;DR

This document presents the proceedings of the Third Workshop on Natural Language Processing for Indigenous Languages of the Americas (AmericasNLP 2023). It highlights the critical need for NLP research on low-resource indigenous languages, showcasing advancements in machine translation, language identification, and corpus development, while emphasizing community involvement and ethical considerations.

What Problem Does This Paper Address?

The vast majority of NLP research and development has historically focused on high-resource languages, leaving a significant gap in the application of these technologies to the world's thousands of low-resource languages, particularly indigenous languages of the Americas. This exclusion leads to a digital divide, hindering the preservation, revitalization, and accessibility of these languages. AmericasNLP aims to bridge this gap by fostering research, connecting communities, and promoting the development of NLP tools and techniques specifically tailored for these underrepresented linguistic communities. The workshop addresses the challenge of building effective NLP systems for languages with limited data, often lacking standardized orthographies, and facing the risk of endangerment.

Methods

The papers presented at AmericasNLP 2023 explore a variety of methods to address the challenges of low-resource NLP. These include:

  • Neural Machine Translation (NMT) with Active Learning: Techniques to efficiently train NMT models for languages with scarce parallel data, such as using English as a pivot language or leveraging monolingual data.
  • Language Identification: Developing models capable of distinguishing between closely related indigenous languages, even with limited training data.
  • Morphological Modeling: Utilizing formalisms like Finite-State Transducers (FSTs) and Recurrent Neural Networks (RNNs) to model complex morphology found in many indigenous languages.
  • Word Embeddings and Dictionary Construction: Methods for creating useful word representations and dictionaries from limited bilingual resources, including fine-tuning pre-trained models and leveraging existing dictionaries.
  • Community-Driven Development: Emphasizing the importance of involving native speakers and communities in the entire NLP development lifecycle, from data collection and annotation to evaluation and application design.
  • Data Augmentation and Synthesis: Strategies to increase the amount of training data, such as back-translation and leveraging diverse data sources.

Datasets

The workshop features research utilizing and contributing to various datasets, often specifically curated for indigenous languages:

  • Spanish to Mapudungun: A dataset used for NMT research.
  • Brazilian Indigenous Languages: Datasets for language identification tasks.
  • Florentine Codex: An open and extensible machine-readable edition used for corpus development.
  • Maya: Datasets for finite-state language technology.
  • Lushootseed: Data used for morphological modeling.
  • Bilingual Dictionaries: Used for constructing word embeddings for low-resource languages.
  • Eastern and Southwestern Ojibwe: A small corpus for dialect identification.
  • Wayunaiki and Spanish: Data for enriching NMT systems.
  • Inuktitut: Data for Named Entity Recognition.
  • Spanish-Mazatec and Spanish-Mixtec: Parallel corpora for translation.
  • Highland Puebla Nahuatl: Data for morphological analysis.
  • Akuzipik-English: Data for community consultation and dictionary development.
  • Low-resource languages in general: Many papers focus on leveraging existing, albeit small, datasets or creating new ones.
  • AmericasNLP 2023 Shared Task Datasets: Specific datasets were created for the shared task, focusing on machine translation into indigenous languages.

Key Insights and Ideas

  • The "200 Language Barrier" is Real and Solvable: Angela Fan's keynote highlights the significant challenge of scaling NLP to a vast number of languages, but also presents a vision and progress towards breaking these barriers with human-centered approaches.
  • Community is Paramount: Multiple papers underscore that successful NLP for indigenous languages cannot be done to communities but with them. This involves deep collaboration, respecting linguistic sovereignty, and ensuring that developed tools benefit the speakers directly.
  • Leveraging Existing Resources is Crucial: For extremely low-resource languages, creative use of existing dictionaries, related languages, and even monolingual data is essential for building functional NLP systems.
  • Ethical Considerations are Non-Negotiable: The development of NLP for indigenous languages must prioritize safety, quality, and the ethical implications of data usage and model deployment, especially concerning sensitive information and potential misuse.
  • Formalisms Still Matter: While neural methods dominate, traditional linguistic formalisms like FSTs remain valuable tools for modeling complex linguistic phenomena in low-resource settings, often complementing neural approaches.

Key Results

  • Improved Machine Translation: Several papers report significant improvements in BLEU scores for machine translation between English and indigenous languages like Mapudungun, Spanish-Mazatec, and Spanish-Mixtec, often outperforming previous state-of-the-art. For instance, one system achieved a 44% BLEU improvement for low-resource languages.
  • Effective Language Identification: Models were developed that can accurately identify indigenous languages, even when they are closely related, demonstrating the feasibility of this task for low-resource scenarios.
  • Functional NLP Tools: Research has led to the development of functional tools such as morphological analyzers, named entity recognition systems, and dictionaries, which are crucial for language documentation and revitalization.
  • Successful Shared Task Participation: The AmericasNLP 2023 Shared Task saw numerous submissions demonstrating progress in machine translation for indigenous languages, with various approaches yielding competitive results.

Limitations and Future Work

  • Data Scarcity Remains a Core Challenge: Despite advancements, the fundamental limitation for many indigenous languages is the sheer lack of digital data, which hinders the development of robust NLP models.
  • Evaluation Difficulties: Evaluating NLP systems for low-resource languages is challenging due to the lack of standardized evaluation benchmarks and the need for human evaluation by native speakers, which can be resource-intensive.
  • Generalization Across Languages: Models trained on one low-resource language may not generalize well to others, necessitating language-specific adaptations and research.
  • Ethical Frameworks Need Further Development: While acknowledged, the practical implementation of ethical guidelines and community-centered development requires ongoing refinement and standardization.
  • Future Work: The papers suggest continued research in areas such as developing more efficient unsupervised and semi-supervised learning methods, creating larger and more diverse datasets, building robust evaluation metrics, and fostering stronger collaborations between NLP researchers and indigenous language communities. The goal of creating a truly universal translation system remains a long-term aspiration.

Summary updated

BibTeX
@inproceedings{mager2023proceedings,
  author = {Manuel Mager and Abteen Ebrahimi and Arturo Oncevay and Enora Rice and Shruti Rijhwani and Alexis Palmer and Katharina Kann},
  title = {{Proceedings of the Workshop on Natural Language Processing for Indigenous Languages of the Americas (AmericasNLP)}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
}
Figure