Apurinã Universal Dependencies Treebank

Jack Rueter, Marília Fernanda Pereira de Freitas, Sidney Da Silva Facundes, Mika Hämäläinen, Niko PartanenNAACL 2021

TL;DR

This paper introduces the first Universal Dependencies (UD) treebank for Apurinã, an endangered Amazonian language. This resource is crucial for advancing NLP research on low-resource languages, enabling the development of tools like parsers and POS taggers, and supporting linguistic research by capturing Apurinã's rich morphology within a standardized framework.

What Problem Does This Paper Address?

Developing NLP tools for endangered languages presents a significant challenge due to the scarcity of annotated data. The Universal Dependencies (UD) project aims to create syntactically annotated corpora for many languages, fostering cross-linguistic consistency. However, many endangered languages, like Apurinã, lack such resources, hindering their inclusion in NLP research and computational linguistics. This paper addresses this gap by constructing the first UD treebank for Apurinã, a language spoken by around 2,000 native speakers and classified as endangered. The process also serves as an opportunity to develop finite-state descriptions and transfer existing open-source infrastructure to this context.

Methods

The authors constructed the Apurinã UD treebank by applying the Universal Dependencies framework to existing fieldwork data. This involved adapting 14 parts-of-speech and seven augmented or new features, some of which are unique to Apurinã. The process leveraged the open-source GiellaLT infrastructure, which provides a finite-state approach to linguistic description and facilitates the transfer of practices from morphologically complex languages like those in the Uralic family. Key challenges included aligning Apurinã's rich morphology with UD conventions, particularly for features like case, number, person, and gender, which required extensions beyond standard UD practices. For instance, Apurinã's case system, with its oblique markers, and its complex possessive constructions necessitated careful feature design. The lemmatization process also had to account for orthographic variations and ongoing language evolution.

Datasets

The paper introduces the Apurinã Universal Dependencies Treebank as a newly created benchmark. The initial release contains 76 fully annotated sentences, comprising 574 tokens and 454 words. The source materials for this treebank were derived from fieldwork practices, with the understanding that not all tokens were initially annotated with the same level of detail.

Key Insights and Ideas

  • UD as a bridge for endangered languages: The creation of this treebank demonstrates how the UD framework can be adapted to capture the linguistic complexity of endangered languages, making them accessible for computational linguistics.
  • Morphological richness and UD adaptation: Apurinã's rich morphology, particularly in case and possession, required novel feature extensions within the UD framework, highlighting the flexibility needed for diverse linguistic structures.
  • Finite-state methods for endangered languages: The use of GiellaLT infrastructure shows the potential of finite-state approaches in building resources for languages with complex morphology, facilitating transferability and consistency.
  • Ongoing annotation practices: The paper acknowledges that establishing regular annotation practices for the entire treebank is an ongoing project, reflecting the iterative nature of linguistic resource development.

Key Results

The Apurinã UD treebank, in its first release, contains 76 annotated sentences. The most salient features observed are Case (101 instances), Gender (96), and Number (73), with the newly introduced Gender[obj] also being well-attested (47 instances). The most prominent parts-of-speech are NOUN (170) and VERB (137), followed by PRON (59) and ADV (39). Dependency relations show nsubj (83) as the most frequent, enabled by extensive use of the conj relation. Language-specific dependency relations include lmod, neg, poss, relcl, tcl, and tmod.

Limitations and Future Work

The authors acknowledge that due to the dataset's size and orientation, some features of the Apurinã language have been neglected. A significant future challenge is to adapt recent studies on noun incorporation for UD, specifically for Apurinã's noun and classifier incorporation. The primary future goal is to expand the treebank to a size suitable for splitting into train, test, and dev portions. This expansion is linked to resource availability, with a plan to connect treebank sentences more closely to existing audio and video recordings and multimodal resources, which are invaluable for language documentation.

Summary updated

BibTeX
@inproceedings{rueter2021apurina,
  author = {Jack Rueter and Marília Fernanda Pereira de Freitas and Sidney Da Silva Facundes and Mika Hämäläinen and Niko Partanen},
  title = {{Apurinã Universal Dependencies Treebank}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2021},
}
Figure