A Relational Representation of Modification

Samuel BayerAAAI 1986

Introduction

In the development of natural language processing (NLP) systems during the mid-1980s, a persistent challenge was the accurate representation of linguistic modifiers—words such as "long," "distance," or "at." While contemporary systems were adept at handling simple properties, they often struggled with abstract relations that spanned multiple domains. This paper presents a unified framework for handling these modifiers by treating them as full-fledged relational entities rather than simple attribute-value pairs. This approach, implemented in the KING KONG parser, leverages relational database principles and Jackendoff’s thematic structures to generalize meanings across spatial, temporal, and possessive contexts.

The Limitations of Attribute-Value Models

Prior to this work, most NLP systems utilized attribute-value representations (such as KL-ONE’s "ROLES"). The authors identify two primary failures in this paradigm:

  1. The N-ary Relation Problem: Attribute-value systems are inherently binary, linking one object to one value. This structure becomes cumbersome when representing three-place relations. For example, the statement "Dresden is 100 miles from Hahn" involves a distance relation between two distinct locations and a specific value, a structure that is difficult to model without redundant or clunky logic in a binary system.
  2. Linguistic Inconsistency: Natural language allows for significant syntactic flexibility. A single semantic relationship can be expressed as "The runway is long," "the long runway," or "the length of the runway." Traditional systems often required separate underlying logic for each syntactic variation, failing to recognize them as expressions of the same semantic core.

The KING KONG Architecture

To resolve these issues, the authors proposed a three-pronged strategy within the KING KONG parser:

1. Relational Representation

Instead of treating a relation as a mere "link" between two nodes, the relation itself is treated as a node (a first-class citizen). This allows for "uniform access," meaning the system can query the value of an attribute or the object bearing that attribute with equal computational ease. By moving away from "links," the system avoids the ontological divisiveness of treating binary and n-ary relations as fundamentally different structures.

2. Argument Structure Shorthand

The authors developed a linguistic shorthand to map various syntactic forms to a single semantic structure. This is achieved through two primary markers:

  • POSS-OBJ: Identifies the object being measured (the "Theme").
  • PRED-MOD: Identifies the value or modifier of that measurement.

3. Thematic Generalization (Types and Families)

Drawing on the work of Ray Jackendoff and Jeffrey Gruber, the framework categorizes relations into Types (e.g., EXTENT, POSITION) and Families (e.g., SPACE, TIME, POSSESSION). A word like "long" designates an EXTENT type, but its specific meaning is derived from the "Family" of its argument. For instance, a "long runway" is interpreted within the SPACE family, while a "long meeting" is interpreted within the TIME family.

Cross-Domain Logic and Spatial Metaphor

A key insight of the paper is the use of spatial metaphor as a logical tool. The authors argue that much of human language is inherently metaphorical, with spatial logic serving as a template for other domains. By defining a core logic for "Space" and then parameterizing it, the KING KONG system can understand "Time" or "Possession" using the same underlying code.

In the Type-Family matrix, some words (like "length") only specify a Type (EXTENT). The system then examines the object (e.g., a runway) to determine the Family (SPACE). If multiple relations exist at that intersection—such as Length, Width, or Altitude—the system utilizes a "default" relation to resolve the ambiguity and select the most likely intended meaning.

Implementation and Results

The system's efficacy was demonstrated through Lisp-like definitions within the KING KONG parser. In qualitative tests, the system successfully resolved complex queries such as "How long is the runway?" By identifying "runway" as an OBJECT in the SPACE family and "long" as an EXTENT type, the parser correctly selected the LENGTH relation as the default for that intersection.

This relational approach offers significant portability. Because the framework focuses on data access rather than rigid data organization, the natural language interface can be more easily ported to different expert systems and databases without requiring a total overhaul of the underlying logic.

Limitations and Future Work

While the framework provides a robust foundation, the authors noted several areas for expansion:

  • Granularity: The argument shorthand requires finer linguistic nuance to handle more complex sentences.
  • Path Functions: Future iterations aim to implement Jackendoff’s distinctions between different types of paths, such as bounded paths versus directions.
  • Morphological Complexity: The authors suggest that suffixes like -ness and -ly could eventually be handled via morphological rules that preserve underlying semantics.
  • Ontological Scope: While effective for OBJECT and EVENT classes, extending the system to other ontological categories remains a primary goal for future research.

Summary updated

BibTeX
@inproceedings{bayer1986relational,
  author = {Samuel Bayer},
  title = {{A Relational Representation of Modification}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure