A High-level Morphological Description Language Exploiting Inflectional Paradigms

Peter Anick, Suzanne ArtemieffCOLING 1992

TL;DR

This paper introduces a new language, PDL, for describing word morphology that leverages paradigm inheritance, mirroring how grammar books organize language. This approach allows for more concise and natural descriptions of inflectional patterns, especially for complex languages like French and German, by sharing common rules across related word classes.

What Problem Does This Paper Address?

Describing the morphology of languages—how words change form to indicate grammatical function (like tense, number, or case)—is a complex task for computational linguists. Traditional approaches often focus on individual rules for stem changes and affixes, but struggle to efficiently capture the shared patterns and regularities between groups of words (paradigms). Existing computational morphology systems often overlook the organizational principles found in traditional grammar books, which group words into paradigms (e.g., verb conjugations). This leads to either overly complex rule sets or a failure to exploit linguistic generalizations, making it difficult to build robust and efficient morphological analyzers, especially for highly inflected languages.

Methods

The paper proposes a Paradigm Description Language (PDL) with three core components:

  • Form Rules: These define how word forms are constructed, distinguishing between surface forms (what appears in text), lexical forms (stored in a lexicon), and intermediate forms (useful for describing steps in word formation). These rules are restricted to basic stem-affix concatenation and removal.
  • Paradigm Inheritance Hierarchy: This is the central innovation. PDL allows paradigms to be defined based on other paradigms, inheriting their rules by default. This "defeasible inheritance" means a new paradigm can add new rules, override existing ones, or delete inherited ones. This mirrors how grammar books describe regular and irregular verbs, where irregular verbs often share many rules with regular ones but have specific exceptions. Abstract paradigms can also be defined to capture commonalities across multiple concrete paradigms.
  • Orthographic Rules: These handle spelling changes that occur during word formation, independent of specific form rules. They can be applied based on phonological or orthographic context and can be limited to specific paradigms or forms.

The language is designed to be compilable into efficient run-time data structures (non-deterministic transition networks) for fast morphological analysis and generation, while minimizing disk access to a secondary storage lexicon.

Datasets

Not discussed in the paper. The authors mention building representations for English, French, and German, and beginning work on Spanish, but do not detail specific datasets used for development or evaluation.

Key Insights and Ideas

  • Grammar Book Analogy: The core idea is that linguistic descriptions in grammar books, organized by paradigms and inheritance, are a good model for computational morphology. PDL aims to capture this descriptive power.
  • Power of Inheritance: Paradigm inheritance is crucial for concisely describing inflectional behavior. It allows for abstracting common rules, reducing redundancy, and handling subregularities (patterns that apply to a subset of a paradigm).
  • Intermediate Forms: The concept of "intermediate forms" is key. These are not necessarily surface forms but represent crucial steps in word formation. Their inheritance allows for sharing complex derivation paths, even if the final surface forms differ.
  • Lexical Override: The ability for lexical entries to "override" or "augment" paradigm rules is essential for handling truly irregular words without creating an excessive number of unique paradigms. This balances generalization with the need for specific exceptions.
  • Separation of Concerns: Separating orthographic rules from form construction rules allows for more general and concise descriptions of spelling changes that apply across different word classes.

Key Results

The paper demonstrates the effectiveness of PDL by showing how it can concisely encode the complex inflectional paradigms for French verbs. They present a hierarchy of six paradigms that capture the regularities and irregularities of French verbs, reducing the need for a large number of individual rules or paradigms. The authors state that their approach supports efficient analysis and generation and minimizes lexicon access.

Limitations and Future Work

  • German Compounding: German compounding is currently not fully supported by PDL's form rules, which are limited to a single stem component. The authors expect most compounds to be stored directly in the lexicon but are exploring heuristics for analyzing them.
  • Multiple Inheritance: While PDL supports single inheritance, the authors acknowledge that some words might appear to derive variants from multiple paradigms. They have opted for simplicity by using extra lexical stems to accommodate such cases rather than implementing multiple inheritance.
  • Future Work: The authors plan to continue "tuning" existing rulesets, expand coverage to more European languages, and integrate PDL with derivational morphology and compounding systems. They also mention the need for "acquisition-only" paradigms to help generate canonical forms for unknown words during lexicon acquisition.

Summary updated

BibTeX
@inproceedings{anick1992high,
  author = {Peter Anick and Suzanne Artemieff},
  title = {{A High-level Morphological Description Language Exploiting Inflectional Paradigms}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1992},
}
Figure