Encoding Lexicalized Tree Adjoining Grammars with a Nonmonotonic Inheritance Hierarchy

Roger Evans, Gerald Gazdar, David Weir

TL;DR

This paper presents a novel approach to encoding Lexicalized Tree Adjoining Grammars (LTAG) using the DATR language, which leverages a nonmonotonic inheritance hierarchy. This method significantly reduces redundancy in LTAG lexicons by treating lexical rules as covariation constraints within feature structures, offering a more compact and efficient representation.

What Problem Does This Paper Address?

Lexicalized Tree Adjoining Grammars (LTAG) are powerful formalisms for natural language syntax, but they suffer from significant redundancy in their lexicons. As LTAG grammars grow in size (e.g., the XTAG system with over 100,000 lexemes), representing the vast number of elementary trees and their relationships efficiently becomes a major challenge. Traditional LTAG lexicons often repeat similar structural information across many entries, lacking a mechanism to capture generalizations. Previous attempts to address this involved hierarchical lexicons and lexical rules, but this paper argues for a more integrated approach using an existing, general-purpose lexical knowledge representation language (LKRL).

Methods

The core of the paper's contribution lies in its use of the DATR language to represent LTAG lexicons as a nonmonotonic inheritance hierarchy. Key methodological aspects include:

  • Bottom-up Feature Encoding: LTAG trees are encoded from their lexical anchors upwards using a variant of Kilbury's bottom-up encoding. This means tree structures are described relative to a distinguished leaf node using binary relations like parent, left, and right. These relations are then embedded as features within DATR's feature structures.
  • Nonmonotonic Inheritance Hierarchy: DATR's nonmonotonic inheritance allows for default specifications to be overridden by more specific ones. This is crucial for building up complex lexical entries from abstract definitions.
  • Lexical Rules as Covariation Constraints: Instead of separate mechanisms for lexical rules, this paper treats them as covariation constraints within DATR's feature structures. This means that transformations like passive voice or dative alternation are expressed by defining how feature structures change, leveraging DATR's existing feature manipulation capabilities.
  • Integration of Tree Structure and Lexical Information: The paper embeds the LTAG tree structure directly into the feature structure. This allows for a unified representation where syntactic information and lexical properties are handled within the same framework.

Datasets

Not discussed in the paper. The paper focuses on the formal representation of LTAG lexicons and does not present empirical evaluation on specific datasets.

Key Insights and Ideas

  • Unified Representation: The most significant insight is that LTAG's radically lexicalist nature, where all grammatical relations are lexically represented, makes it amenable to a unified representation within a single LKRL like DATR. This avoids the need for separate formalisms for tree structure, lexical properties, and lexical rules.
  • Efficiency through Inheritance: By organizing the lexicon hierarchically with DATR, the paper demonstrates how to eliminate considerable redundancy. Abstract verb classes (e.g., VERB, VERB+NP, VERB+NP+PP) are defined, and specific lexemes (like Give) inherit their properties, only specifying what is unique to them.
  • Lexical Rules as First-Class Citizens: Treating lexical rules as covariation constraints within the inheritance hierarchy is a powerful idea. This means that rules are not external operations but integral parts of the lexicon, subject to the same inheritance and exception mechanisms as other lexical properties.
  • Flexibility and Extensibility: The DATR approach allows for easy extension to other parts of speech and grammatical constructions. The paper illustrates this with examples for auxiliary verbs and shows how different types of rules (e.g., passive, dative, wh-questions) can be encoded.

Key Results

The paper demonstrates the feasibility and benefits of encoding LTAG lexicons using DATR and a nonmonotonic inheritance hierarchy. While no quantitative results are presented, the key qualitative findings are:

  • Compact Lexicon Representation: The proposed method significantly reduces redundancy compared to flat LTAG lexicons.
  • Integrated Lexical Rules: Lexical rules are seamlessly integrated into the inheritance hierarchy, allowing for exceptions and generalizations in a natural way.
  • Unified Framework: The approach provides a single, consistent framework for representing LTAG trees, lexical properties, and lexical rules, leveraging the power of DATR.

Limitations and Future Work

  • Complexity of Full Implementation: The paper acknowledges that the full DATR fragment for the XTAG system is more complex than the simplified examples presented, particularly concerning the handling of unbounded dependency constructions and the precise definition of tree structures.
  • Rule Application: The current DATR encoding of lexical rules primarily supports single application. The authors note that multiple applications could be supported by creating distinct rule instances, but this would require different rule names.
  • Comparison with Other Approaches: While the paper compares its approach to Vijay-Shanker & Schabes (1992) and Becker (1993; 1994), it suggests that its unified, nonmonotonic approach offers advantages in terms of integration and handling of exceptions.
  • Future Work: The authors imply that further work could explore the application of this framework to a wider range of constructions and parts of speech, and potentially refine the mechanisms for handling multiple rule applications. The paper also points to the potential for DATR to integrate with other levels of linguistic description (phonology, semantics, etc.).

Summary updated

BibTeX
@inproceedings{evans1995encoding,
  author = {Roger Evans and Gerald Gazdar and David Weir},
  title = {{Encoding Lexicalized Tree Adjoining Grammars with a Nonmonotonic Inheritance Hierarchy}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1995},
  doi = {10.3115/981658.981669},
}
Figure