Developing a Part-Of-Speech tagger for te reo Māori

Aoife Finn, Peter-Lucas Jones, Keoni Mahelona, Suzanne Duncan, Gianna LeoniACL 2022

TL;DR

This paper introduces Whakairo Kupu, a novel Part-of-Speech (POS) tagger specifically designed for te reo Māori, the Indigenous language of Aotearoa, New Zealand. By developing a custom tagset that reflects Māori linguistic concepts and speaker intuition, the authors achieve high precision and recall, overcoming limitations of existing universal POS tagsets.

What Problem Does This Paper Address?

Developing accurate NLP tools for under-resourced languages, especially Indigenous ones, presents significant challenges. Existing universal POS tagsets, often based on European language structures, fail to adequately capture the unique grammatical nuances and conceptualizations of languages like te reo Māori. This paper addresses the critical need for a POS tagger that is linguistically appropriate for te reo Māori, ensuring that its grammatical categories and tagging conventions reflect the language's own worldview and structure, rather than imposing external frameworks. This is crucial for the re-Indigenization and decolonization of digital language resources.

Methods

The core of the methodology involved creating a bespoke POS tagset for te reo Māori. This was achieved through extensive collaboration with highly-proficient Māori speakers and linguists. The process involved:

  • Adapting Universal Dependencies (UD): The project started by basing its tagset and guidelines on the UD conventions, a widely used standard.
  • Customizing Existing Tags: Many existing UD tags were re-purposed or had their definitions altered to better suit Māori.
  • Creating New Tags: Entirely new POS labels were introduced where existing categories were insufficient or inappropriate. This included labels like MOD (modifier) and ADPRON (adposition-pronoun) to capture specific Māori word classes.
  • Speaker-Centric Approach: The development process was driven by asking non-leading questions to Māori speakers to understand how they conceptualize their language, avoiding academic or pedagogical biases.
  • Iterative Refinement: The tagset and guidelines were designed to be "evergreen," meaning they could evolve based on speaker feedback. An automation system was put in place to retag corpora when changes occurred.

Datasets

The paper mentions that the POS tagger was trained and evaluated on datasets containing over 40,000 tokens, covering many genres and being constantly expanded. No specific benchmark datasets were introduced, but the developed annotated datasets serve as a new resource for te reo Māori NLP.

Key Insights and Ideas

  • Linguistic Appropriateness is Paramount: Imposing grammatical categories from one language family onto another is not only inaccurate but also risks homogenizing and devaluing Indigenous languages. A language's structure and concepts must be understood from the perspective of its speakers.
  • "Particles" are a Complex Word Class: Te reo Māori "particles" are highly versatile and do not neatly fit into traditional grammatical categories like "particle" in other languages, which often serve as a catch-all for words that don't fit elsewhere. This necessitated new tags.
  • Code-Switching Requires Nuance: Simply labeling foreign words with a generic "X" tag fails to capture the complex bi-cultural reality of places like Aotearoa. Distinguishing between languages like English and other Polynesian languages requires specific labels (PAKEHA for English, MOANANUI for other Polynesian languages).
  • Māori Worldview Matters: The tagset development actively sought to reflect a Māori worldview, avoiding terms like "dialect" when not appropriate and focusing on concepts like "kupu mino" (loanwords) and "te reo ā-kāinga" (local language).

Key Results

The paper reports that the Whakairo Kupu POS tagger achieved a precision of 92.5% and a recall of 93.1% in its most recent model. This represents a significant improvement from the very first model, which had 86.3% precision and 48.3% recall, highlighting the effectiveness of the custom tagset and speaker-driven development.

Limitations and Future Work

  • Limitations: The authors acknowledge that while they adapted UD, they did not create new tags for every instance where UD was applicable, indicating that some UD tags were still used. They also note that for code-switching scenarios, a direct application of UD's approach (tagging both languages) would not be feasible for their small organization due to resource constraints and the need for permission from speakers of other languages.
  • Future Work: The paper mentions that Whakairo Kupu is already being used to build a grammar checker and a Named Entity Recognition (NER) tagger for te reo Māori. The approach of developing language-specific tagsets and involving native speakers is presented as a model for other Indigenous languages. The authors also plan to expand the feature layer to include features relevant to Māori, such as "kupu mino" and "te reo ā-kāinga."

Summary updated

BibTeX
@inproceedings{finn2022developing,
  author = {Aoife Finn and Peter-Lucas Jones and Keoni Mahelona and Suzanne Duncan and Gianna Leoni},
  title = {{Developing a Part-Of-Speech tagger for te reo Māori}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2022},
}
Figure