Nonmonotonlc Inheritance and Generic Reflexives

AAAI 1988

TL;DR

This paper introduces a novel nonmonotonic inheritance system to model generic reflexive statements (e.g., "Elephants love themselves") without relying on explicit variables. This approach offers a more linguistically plausible representation of reflexives, bridging the gap between logical formalisms and natural language structure.

What Problem Does This Paper Address?

Traditional approaches to formalizing generic reflexive statements, like "Elephants love themselves," often rely on predicate logic with variables to mark coreferentiality. However, human languages don't typically use such explicit variables at the surface level. This discrepancy poses a challenge for AI systems aiming to understand natural language. Furthermore, first-order logic struggles with generic statements like "Elephants are gray," necessitating more expressive formalisms. This paper aims to provide an alternative semantic theory for reflexives that aligns better with linguistic intuition and leverages the power of nonmonotonic knowledge representation.

Methods

The core of the paper is an extension of Touretzky's nonmonotonic inheritance system. The authors introduce a graphical notation for representing knowledge, including IS-A, IS-NOT-A, and relational links. They define "taxonomic paths" and "relational paths" to represent chains of inheritance. The key innovation lies in extending this system to handle reflexive and irreflexive statements. This is achieved by introducing new rules for inheriting "reflexive paths" and "irreflexive paths," which can be either explicit (directly stated) or implicit (derived from ordinary relational paths that "double back" on themselves). Crucially, the system incorporates mechanisms for contradiction and preemption, allowing for exceptions and overriding of inherited properties, which are essential for handling generic statements.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Eliminating Variables for Reflexives: The system demonstrates that it's possible to represent reflexive statements without resorting to explicit variables, making the formalism more akin to natural language structure.
  • Implicit Reflexives from Generic Relations: Generic statements about relations (e.g., "Herbivores like gray things") can implicitly generate reflexive statements (e.g., "Elephants like themselves") through the inheritance mechanism, even without explicit "self-love" assertions.
  • Nonmonotonicity is Crucial: The ability to handle exceptions and preemption (e.g., "Royal elephants don't like themselves" overriding a general "Elephants like themselves" rule) is fundamental to accurately modeling generic statements.
  • Path-Based Representation: The use of inheritance paths provides a natural mapping to linguistic structures, where nodes represent entities and links represent relations, offering a more intuitive representation than traditional predicate logic.

Key Results

The paper presents a formal system capable of deriving new generic reflexive statements and statements about individuals. It demonstrates how implicit reflexive paths can be generated from ordinary relational paths through nonmonotonic inheritance. The system successfully handles cases where generic statements are overridden by more specific exceptions, such as "Royal elephants don't like themselves."

Limitations and Future Work

The authors acknowledge that they have not yet proven the constructibility (or at least the existence) of extensions for their system, although they are confident it's possible. They also note that their account of "other" is not exhaustive, particularly regarding existential interpretations and the nuances of its use in English. Future work could explore these linguistic subtleties further and potentially translate the system into other formalisms like default logic.

Summary updated

BibTeX
@inproceedings{1988nonmonotonlc,
  title = {{Nonmonotonlc Inheritance and Generic Reflexives}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1988},
}
Figure