Representation of Yine [Arawak] Morphology by Finite State Transducer Formalism

Adriano Ingunza Torres, John Miller, Arturo Oncevay, Roberto Zariquiey BiondiNAACL 2021

TL;DR

This paper presents a Finite State Transducer (FST) based morphological analyzer for Yine, a critically endangered polysynthetic language. The authors tackle the complexity of Yine's morphology, particularly its rich verbal and nominal structures, to build a tool that aids in language documentation and preservation, with potential applications in NLP tasks like spell checking.

What Problem Does This Paper Address?

Polysynthetic languages, like Yine, are characterized by their complex morphology, where a single word can express the meaning of an entire sentence. This complexity poses a significant challenge for computational linguistic tools, especially for low-resource languages where documentation and existing NLP resources are scarce. The UNESCO classifies Yine as "definitely endangered," highlighting the urgent need for tools to support its continued vitality. While FSTs are a well-established formalism for morphology, applying them to highly agglutinative and polysynthetic languages like Yine requires novel approaches to handle intricate patterns of affixation, incorporation, and morphophonological processes like vowel deletion and rhotacism. The authors aim to bridge this gap by developing a high-coverage FST morphological analyzer for Yine.

Methods

The core of this work is the development of a Finite State Transducer (FST) morphological analyzer. The authors leverage the Foma library, a tool for constructing FSTs, and follow a "divide and conquer" strategy based on the existing grammar of Yine. They define FST components functionally as lexical, post-lexical, and memory filters, along with lexicons and alterations.

Key methodological aspects include:

  • FST Construction: Building transducers to map between morphological analyses and word forms.
  • Handling Polysynthesis: Developing techniques to manage complex patterns of incorporating open and closed category arguments, and multiple morphemes.
  • Morphophonological Rules: Implementing FSTs to model processes like vowel deletion and rhotacism, including constraints on these processes (e.g., avoiding three-consonant clusters).
  • Possessive Constructions: Modeling alienable and inalienable possession, which involves complex prefixation and suffixation rules, and distinctions based on noun root class.
  • Verbal Morphology: Representing intricate verbal structures, including noun incorporation, aspect, and argument indexing.
  • Lexical Filters and Alterations: Using these FST components to restrict lexical generation, handle over-generation, and implement phonological and morphophonological constraints.

Datasets

The authors utilize several linguistic resources for their analysis and evaluation:

  • Hanson (2010) Grammar: A comprehensive typological grammar of Yine, serving as the primary source for morphological rules and analysis.
  • Yine-Spanish/Spanish-Yine Dictionary (Wise, 1986): Used for lexical information.
  • Theoretical Guide (Zapata et al., 2017): Provided theoretical background.
  • Yine Corpus (Bustamante et al., 2020): A corpus of Yine text used for evaluation purposes.

No new datasets were introduced in this paper.

Key Insights and Ideas

  • FSTs as a Powerful Tool for Endangered Languages: The paper demonstrates that FSTs, despite the complexity of polysynthetic languages, are a viable and powerful formalism for documenting and preserving endangered languages like Yine.
  • The Challenge of "Solved Problems": While morphology is often considered a "solved problem" in computational linguistics, the authors highlight that representing highly complex, low-resource languages like Yine is far from trivial and requires significant effort and specialized techniques.
  • Interplay of Lexical Specification and Phonological Constraints: The analysis of vowel deletion in Yine reveals that while some deletions might be lexically specified, phonological constraints (like avoiding consonant clusters) play a crucial role in determining the actual surface form.
  • The Importance of Detailed Grammatical Descriptions: The reliance on Hanson's (2010) grammar underscores the critical role of thorough linguistic documentation in building computational models for complex languages.
  • The Need for Robust NLP Tools for Language Vitality: The ultimate goal of creating an FST analyzer for spell-checking and translation emphasizes the direct link between computational linguistics and the preservation and revitalization of endangered languages.

Key Results

The paper reports on the development of a partial FST morphological analyzer for Yine, covering nominal and verbal constructions, including possessive constructions and verbal predicates.

  • Coverage: A preliminary evaluation on a sample of 110 nouns and 110 verbs (from a corpus by Bustamante et al., 2020) showed approximately 15% coverage. This is acknowledged as a low figure, interpreted as a measure of the effort still required.
  • Error Analysis: The evaluation identified several reasons for lack of coverage, including:
  • Morphemes not present in the FST vocabulary.
  • Incorrect root entries or orthographic mismatches.
  • Issues with elision and rhotacism alterations.
  • Challenges with non-verbal predicates and category changes.
  • Qualitative Findings: The authors successfully modeled complex phenomena such as:
  • Possessive constructions with alienable/inalienable distinctions.
  • Vowel deletion and rhotacism, including their constraints.
  • Noun incorporation into verbal predicates.
  • Argument indexing for subjects and objects.

Limitations and Future Work

The authors explicitly state that their analyzer is a work in progress and that the implementation is incomplete.

Limitations:

  • Partial Coverage: The current analyzer has low coverage, indicating significant work remains.
  • Incomplete Grammar Description: The authors acknowledge that Hanson's (2010) grammar is not exhaustive, and further study of Yine's morphological patterns is needed.
  • Limited Community Support for Foma: While Foma is used, its limited community support is noted.

Future Work:

  • Increasing Coverage: The primary goal is to achieve high coverage for the FST morphological analyzer.
  • Developing More Components: Expanding the analyzer to include more word categories and functions, such as nominalizers and verbalizers.
  • Testing with External Data: Continued testing with real-world data to identify and fix deficiencies.
  • NLP Applications: Integrating the analyzer into NLP applications like spell checkers and machine translators to promote language vitality.
  • Further Research on Morphophonological Rules: Deeper investigation into the complex morphophonological rules of Yine.

Summary updated

BibTeX
@inproceedings{torres2021representation,
  author = {Adriano Ingunza Torres and John Miller and Arturo Oncevay and Roberto Zariquiey Biondi},
  title = {{Representation of Yine [Arawak] Morphology by Finite State Transducer Formalism}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2021},
}
Figure