The First Komi-Zyrian Universal Dependencies Treebanks

Niko Partanen, Rogier Blokland, KyungTae Lim, Thierry Poibeau, Michael RießlerEMNLP 2018

TL;DR

This paper introduces the first Universal Dependencies (UD) treebanks for Komi-Zyrian, a critically endangered Uralic language. By leveraging existing language documentation tools and a collaborative approach, they've created valuable resources that not only advance NLP for low-resource languages but also support the crucial task of documenting endangered linguistic heritage.

What Problem Does This Paper Address?

Documenting and analyzing endangered languages presents a significant challenge. Traditional methods often lack the scalability and comparability needed for modern computational linguistics. Furthermore, existing NLP resources and frameworks, like Universal Dependencies (UD), are heavily biased towards well-resourced languages, leaving many endangered languages underrepresented. This gap hinders both linguistic research and the development of NLP tools that could aid in their preservation. The authors tackle this by creating UD treebanks for Komi-Zyrian, a language with around 160,000 speakers but threatened by Russian. They aim to bridge the divide between traditional language documentation practices and the needs of computational linguistics, making Komi-Zyrian data more accessible and usable for NLP tasks.

Methods

The creation of the Komi-Zyrian UD treebanks involved a multi-stage process, integrating existing linguistic tools with the UD framework. The core of their methodology relies on Giellatekno's open infrastructure, which provides a mature morphosyntactic analyzer for Komi. This system, built on FST Morphology and Constraint Grammar, performs tokenization, morphological analysis, and rule-based disambiguation.

The process then involved:

  1. Syntactic Analysis: Sending Komi sentences to the Giellatekno analyzer.
  2. Ambiguity Resolution: Manually resolving ambiguities identified by the analyzer.
  3. Manual Annotation: Adding missing syntactic relations using the UD Annotatrix tool. This tool significantly improved annotation speed and consistency, marking a mid-point in the project.
  4. Tag Conversion: Automatically converting the analyzer's XPOS-tags to UPOS-tags and adapting morphological features to UD standards using Python scripts.
  5. Manual Correction and Verification: A final step to ensure accuracy and adherence to UD guidelines.

This approach combines automated analysis with crucial human oversight, a necessity for low-resource languages where fully automated parsing is still challenging. They also highlight the potential for integrating these NLP tools into existing language documentation workflows.

Datasets

The paper introduces two new Komi-Zyrian treebanks:

  • Lattice Treebank: Primarily based on written Komi-Zyrian texts, including digitized books, newspaper articles, and non-fiction works. The data is sourced from the FU-Lab in Syktyvkar, Russia, which has digitized a vast collection of Komi publications.
  • IKDP Treebank: Focuses on spoken Komi-Zyrian data, derived from the IKDP project (Language Documentation Corpus of Komi). This corpus contains transcribed spoken language, including dialectal texts.

Both treebanks are included in the Universal Dependencies 2.2 release. The authors emphasize that the Lattice treebank exclusively uses openly available data to ensure broad reusability.

Key Insights and Ideas

  • Bridging Documentation and NLP: The project demonstrates a successful model for integrating traditional language documentation efforts with modern NLP frameworks like UD. This is crucial for endangered languages where existing documentation is often the primary data source.
  • UD as a Unifying Framework: Universal Dependencies offers a valuable, cross-linguistically comparable annotation scheme that can make data from endangered languages more accessible to a wider research community and facilitate multilingual NLP applications.
  • The Value of Collaboration: The creation of these treebanks was a collaborative effort, highlighting the importance of partnerships between linguists, computational linguists, and language documentation specialists.
  • Challenges in Case Systems: Komi-Zyrian's complex case system, with nuances like approximative and egressive cases, presents theoretical challenges for UD annotation. The paper raises questions about how to best represent such complex grammatical features in a standardized framework.
  • Spoken vs. Written Data: The authors note the need to maintain a clear distinction between written and spoken language representations, especially when dealing with phenomena like code-switching in spoken data.

Key Results

  • The paper presents two Komi-Zyrian UD treebanks: the Lattice treebank with 1389 tokens and the IKDP treebank with 988 tokens.
  • These treebanks have been successfully integrated into the Universal Dependencies 2.2 release.
  • The methodology has proven effective in converting annotations from Giellatekno's scheme into the UD scheme.
  • The UD Annotatrix tool significantly improved annotation speed and consistency.

Limitations and Future Work

  • Size of Treebanks: The current treebanks are relatively small, which has led to them not being split into test and development sets.
  • Annotation Compromises: The authors acknowledge that tagging according to UD principles can be a compromise, potentially not capturing all language-specific particularities. They suggest using the misc column or providing varying degrees of annotation for different parts of the documentary corpus to address this.
  • Future Work:
    • Expanding Treebanks: The goal is to reach 5,000 tokens in both treebanks by the end of 2018 and 20,000 by mid-2019.
  • Incorporating More Genres: Future releases of the Lattice treebank will include more genres, such as newspaper texts and Wikipedia articles.
  • Historical Texts: The IKDP treebank will incorporate individual texts from historical Komi recordings made in the 1950s and 1960s.
  • English Glosses: The authors consider adding English glosses to the misc-field to make the data more accessible to a broader linguistic audience.
  • Testing Parsers: They are interested in testing various approaches (rule-based, statistic-based, hybrid) to replace manual annotation with automated parsers.
  • Case System Annotation: Further work is needed to address the theoretical questions surrounding the annotation of Komi-Zyrian's complex case system within UD.
  • Code-Switching: Verification of Russian annotations in the IKDP treebank to ensure consistency with monolingual Russian treebanks is an ongoing task.

Summary updated

BibTeX
@inproceedings{partanen2018first,
  author = {Niko Partanen and Rogier Blokland and KyungTae Lim and Thierry Poibeau and Michael Rießler},
  title = {{The First Komi-Zyrian Universal Dependencies Treebanks}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2018},
}
Figure