A Finite-State Model of Georgian Verbal Morphology

Olga GurevichNAACL 2006

TL;DR

This paper presents a novel computational model for generating and recognizing Georgian verb conjugations, tackling the language's notoriously complex and non-concatenative verbal morphology. By employing a templatic, word-level analysis combined with finite-state transducers and a default inheritance mechanism, the model offers a viable and extensible approach to capturing Georgian verb structure.

What Problem Does This Paper Address?

Georgian verbal morphology is exceptionally complex, characterized by a large number of inflectional categories, intricate inter-dependencies between elements, and a vast array of regular, semi-regular, and irregular patterns. This complexity makes computational modeling a significant challenge. Traditional approaches often assume morpheme-level compositionality, which proves inadequate for Georgian. The authors aim to develop a computational model that can effectively handle this complexity by moving away from strict compositional rules and instead focusing on a templatic, word-based analysis at various levels of regularity.

Methods

The authors propose a multi-level finite-state transducer (FST) model. The core idea is to compile networks where the output of a lower level serves as input to a higher level. This layered approach breaks down the problem into:

  • Level 1: Lexicon: Stores lexically specific information, including complete word forms for representative verbs of each class, preverbs, and exceptional forms. This level uses two-level FST arcs to map verb stems with morphosyntactic properties to their inflected forms.
  • Level 2: Semi-regular Patterns: Compiles inflectional forms dependent on Level 1 outputs and provides default inflections for regular tense formation. This involves adding suffixes and replacing tense properties.
  • Level 3: Regular Patterns: Affixes regular inflection for object and non-3rd person subject agreement, which are expressed via a combination of prefixal and suffixal markers that attach simultaneously.

The model also incorporates a default inheritance mechanism, allowing for specification of exceptional features at Level 1 and application of default rules at later levels, minimizing manual entry. The implementation uses the xfst program.

Datasets

Not discussed in the paper. The authors mention using data from existing descriptions of Georgian grammar and consulting with native speakers for data.

Key Insights and Ideas

  • Templatic, Word-Level Analysis: The paper argues successfully that a templatic, word-based analysis is more suitable for Georgian verbal morphology than a strictly compositional morpheme-by-morpheme approach.
  • Multi-Level FST Architecture: The layered FST approach effectively decomposes the complex morphology into manageable levels of regularity, from highly lexical to completely regular patterns.
  • Default Inheritance for Exceptions: Using default inheritance significantly reduces the manual effort required by allowing exceptional forms to be specified at the lowest level, with general rules applying elsewhere.
  • Handling Lexical Variation: The model acknowledges and addresses the significant lexical variation in Georgian verbs, including preverb differences and stem alternations, by treating these as lexically specific information at Level 1.

Key Results

The paper presents a functional computational model for generating and recognizing a subset of Georgian verbs. While specific quantitative results (like accuracy percentages) are not detailed, the authors demonstrate the model's operation through examples of verbs like 'paint' and 'open' across the different levels. The model is described as viable, extensible, and capable of capturing generalizations in Georgian verb structure. It is also noted as the only computational model of Georgian verbs currently in active development and available to the non-Georgian academic community.

Limitations and Future Work

  • Limited Scope: The model currently focuses on regular transitive verbs and frequent irregular verbs. The authors plan to expand it to include regular unaccusative, unergative, and indirect verbs, aiming for full inflections for the 200 most frequent Georgian verbs.
  • Generalization to Novel Items: Acknowledged limitation of most finite-state models is their inability to generalize to novel items as humans do. The authors suggest that the model's output could be used to generate training sets for connectionist or statistical models.
  • Lexical Class Handling: While the input to Level 1 contains representatives for each lexical class, straightforward substitution for other verbs within the same class is complicated by variations in preverbs, stem alternations, and suppletion.
  • Future Directions: Expanding the model to cover more verb classes and types, and exploring its use in generating training data for machine learning models.

Summary updated

BibTeX
@inproceedings{gurevich2006finite,
  author = {Olga Gurevich},
  title = {{A Finite-State Model of Georgian Verbal Morphology}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2006},
}
Figure