Generating the XTAG English Grammar Using Metarules

Carlos A. ProloCOLING 2002

TL;DR

This paper presents a method for automatically generating a large, wide-coverage Lexicalized Tree Adjoining Grammar (LTAG) for English. By combining a set of "metarules" (grammar transformation rules) with a structured hierarchy of rule application, the XTAG project successfully generated thousands of verb trees from a much smaller initial set, significantly streamlining grammar development and maintenance.

What Problem Does This Paper Address?

Developing and maintaining large-coverage natural language grammars, particularly Lexicalized Tree Adjoining Grammars (LTAGs), is a significant engineering challenge. LTAGs consist of elementary trees that are combined through operations like adjunction and substitution. As the grammar grows to cover more linguistic phenomena, the sheer number of required elementary trees becomes enormous, making consistent development and maintenance difficult. The XTAG project faced this issue, especially with verb trees, which exhibit a vast diversity of syntactic structures. The paper addresses the need for a more efficient and principled way to generate and manage these large grammars, aiming to capture the regularities in tree structures more concisely than by listing every single tree.

Methods

The core of the paper's methodology lies in the application of metarules, inspired by Becker's work. These metarules are essentially transformation rules that operate on existing trees to generate new ones. The authors combined these metarules with a specific hierarchy of rule application.

Here's a breakdown:

  • Tree Templates: The grammar is built from a set of "tree templates" (elementary trees) that are lexicalized by specific syntactic categories.
  • Metarules: These are rules that describe how to derive related syntactic structures. For example, a metarule can take a declarative sentence tree and generate its passive or wh-extracted counterpart. The paper illustrates this with a metarule for wh-subject extraction.
  • Hierarchy of Rule Application: The authors propose an ordered set of metarules, applied in a "cumulative mode." This means that after applying the first metarule to the initial set of trees, the resulting trees are added to the set, and then the second metarule is applied to this expanded set, and so on. This process is designed to generate entire families of related trees from a single base tree.
  • Feature Equations: Metarules use features for both matching tree structures and for transformation purposes, ensuring that the generated trees are linguistically sound.

Datasets

Not discussed in the paper. The paper focuses on the generation of grammar rules and trees rather than training on a corpus. The "dataset" in this context is the existing XTAG English grammar and the set of metarules themselves.

Key Insights and Ideas

  • Metarules as a Compact Representation: The paper strongly advocates for metarules as an effective way to represent the regularities and relationships between different syntactic structures within an LTAG, leading to a significantly more compact grammar representation.
  • Engineering Efficiency: The approach offers substantial engineering benefits for grammar development and maintenance. Instead of manually creating thousands of similar trees, a smaller set of base trees and metarules can generate them automatically.
  • Ordered Application Suffices: The authors argue that a simple ordering mechanism for metarules is sufficient to manage the generation process and avoid issues of unboundedness, a point of contention in previous work on recursive grammar generation.
  • Separation of Formalism and Control: They propose a clear distinction between metarules as a formal system for tree transformation and a control mechanism that dictates the order of application, suggesting this separation is key to managing complexity.
  • Identifying Grammar Weaknesses: The process of comparing generated trees with an existing grammar highlighted inconsistencies and areas for improvement in the original grammar's design, particularly at the interface between the lexicon and tree templates.

Key Results

  • Massive Tree Generation: The XTAG English grammar's verb trees, which initially comprised 783 trees across 57 subcategorization groups, were successfully generated from a much smaller set of base trees (53 declarative trees) and 21 metarules.
  • Compactness Achieved: The approach demonstrated the ability to generate a large-scale grammar from a compact initial set, significantly reducing the effort required for development and maintenance.
  • Overgeneration and Undergeneration: While the system largely succeeded, it also exhibited some overgeneration (generating trees that are not linguistically valid for certain lexical items) and undergeneration (failing to generate certain idiosyncratic trees). These issues were often linked to the separation of lexicon and grammar rules.

Limitations and Future Work

  • Overgeneration: The authors acknowledge that the metarule system sometimes overgenerates trees, particularly for passive constructions and in families with complex structures or idiomatic expressions. This is partly attributed to the separation of lexicon and grammar rules, suggesting that lexical items could be used to block the selection of certain generated trees.
  • Undergeneration: Some specific tree types (e.g., for "-ed" adjectives, Determiner Gerund trees, and trees involving punctuation) were not generated and need to be added separately.
  • Limitations of Metarule Specification: Current metarules do not allow for specifying negative structural constraints, requiring multiple metarules for certain cases.
  • Feature Equation Complexity: Handling subtle differences in feature equations across the grammar remains a significant challenge, sometimes necessitating separate metarules.
  • Future Work: The authors suggest future work on developing a graph-based interface for metarules that supports partial ordering and distinct treatments for different grammar families. They also plan to explore how lexical items can more directly influence the generation process to mitigate overgeneration.

Summary updated

BibTeX
@inproceedings{prolo2002generating,
  author = {Carlos A. Prolo},
  title = {{Generating the XTAG English Grammar Using Metarules}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2002},
}
Figure