A principle-based hierarchical representation of LTAGs

Marie-Helene CanditoCOLING 1996

TL;DR

This paper introduces a novel, principle-based hierarchical representation for Lexicalized Tree Adjoining Grammars (LTAGs). This approach aims to solve redundancy issues in LTAG development by automatically generating tree families from a compact, linguistically motivated structure, leading to more efficient grammar maintenance and extension.

What Problem Does This Paper Address?

Developing large-scale Lexicalized Tree Adjoining Grammars (LTAGs) is notoriously challenging due to significant redundancy. Hundreds of elementary trees, each with associated feature structures, often repeat information about syntactic structures and well-formedness principles. This redundancy makes grammars difficult to write, update, and maintain, hindering clarity and extensibility. Existing solutions, like inheritance networks and lexical rules, offer some efficiency but lack a principled linguistic basis for the hierarchy and don't fully address the generative aspect of creating new grammar components.

Methods

The authors propose a system that organizes LTAG syntactic descriptions hierarchically, moving away from manual, redundant representations. The core of their approach is a monotonic inheritance network that represents tree schemata. This network is structured along three dimensions:

  1. Canonical subcategorization frame: Defines the basic argument structure of a predicate.
  2. Redistribution of syntactic functions: Captures how argument roles change (e.g., passive voice, dative shift).
  3. Syntactic realizations of functions: Specifies how these functions are realized in the actual sentence structure (e.g., canonical position, extracted position).

Instead of manual rule-writing, the system uses principles of well-formedness (like predicate-function co-occurrence) to automatically generate elementary trees and their families. This generation process involves creating "terminal classes" that inherit properties from their super-classes and then translating these classes into LTAG elementary tree schemata. The system also leverages partial descriptions of trees and a unification principle to manage inheritance and ensure consistency.

Datasets

Not discussed in the paper. The paper focuses on the formalisms and generation process rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • Principle-based Hierarchy: The core idea is that a linguistically motivated hierarchy, guided by well-formedness principles, can inherently capture the relationships between different syntactic structures, thus eliminating redundancy.
  • Monotonic Inheritance: Using monotonic inheritance for syntactic descriptions is argued to be more appropriate than default inheritance for representing tree schemata, leading to greater declarativity and transparency.
  • Automatic Generation: The system's ability to automatically generate tree families based on these principles is a significant step towards efficient LTAG development, allowing for systematic exploration of linguistic phenomena.
  • Separation of Concerns: The three-dimensional hierarchy effectively separates the core predicate argument structure from its various transformations and surface realizations, providing a clearer conceptual organization.
  • Unification for Consistency: The use of unification for partial descriptions and meta-features ensures that inherited properties are consistently integrated, preventing contradictions.

Key Results

The paper presents a formal framework and a system design for generating LTAG elementary trees. While specific quantitative results are not detailed, the authors report on the application of their tool to update and augment the French LTAG. They state that the hierarchy provides a compact and transparent representation, and the automatic generation ensures well-formedness. They also mention adding new phenomena like nominal subject inversion and causative constructions. The tool generated 730 trees for 17 verb families, with a significant portion of these being new compared to the pre-existing grammar.

Limitations and Future Work

  • Technical vs. Linguistic Principles: While the authors advocate for a principle-based approach, they acknowledge that existing solutions (like Becker's) have primarily focused on technical aspects of representation. They aim to bridge this gap by formalizing well-formedness principles.
  • Generative Aspect: The paper highlights that previous solutions lacked a developed generative aspect. Their proposed system directly addresses this by automatically generating tree families.
  • Future Work: The authors suggest several potential applications:
  • Generating grammars with weaker constraints for handling recurrent ill-formed sentences in corpora.
  • Building grammars for other languages, leveraging the French grammar's hierarchy for compatibility and efficiency.
  • Exploring the generation of grammars based on different linguistic theories (e.g., GB perspective).
  • The paper also mentions the need for a formal mechanism to handle the "bounding of meta-rules" (a concept from Becker's work), though their system avoids meta-rules.

Summary updated

BibTeX
@inproceedings{candito1996principle,
  author = {Marie-Helene Candito},
  title = {{A principle-based hierarchical representation of LTAGs}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1996},
}
Figure