Commonsense Metaphysics and Lexical Semantics

Jerry R. Hobbs, William Croft, Todd R. Davies, Douglas Edwards, Kenneth I. LawsACL 1986 · DOI

Abstract

This paper proposes a framework for "commonsense metaphysics"—a set of foundational logical axioms for concepts like time, space, and causality—to bridge the gap between how we use words and how AI reasons about the physical world. By building a rich, minimal core of theories, the authors demonstrate how complex lexical items (like "wear" or "clog") can be formally defined to enable deep automated reasoning in natural language processing.

Problem Statement

In the 1980s, natural language processing faced a significant hurdle: lexical semantics (the study of word meanings) was often disconnected from the "world knowledge" required for true understanding. For example, a system might know that "wear" is a verb, but it wouldn't understand that "wear" implies a loss of material that could eventually lead to a machine's failure. The authors tackle the challenge of encoding the vast, messy "commonsense" knowledge humans use to interpret text, specifically focusing on the TACITUS project for understanding technical manuals about mechanical device failures.

Methodology

The researchers developed a knowledge base of core theories using first-order logic. Their approach is guided by several novel methodological principles:

  • Characterization over Definition: Instead of seeking perfect "if and only if" definitions, they aim for sets of necessary and sufficient conditions.
  • Minimal Structure: Theories are kept abstract (like group theory) to allow the same logic to be applied to metaphorical uses of language (e.g., "high" for physical height vs. numerical values).
  • Dual Ontologies: For complex domains like time, they accommodate multiple viewpoints—treating time both as a continuous "time line" and as a discrete sequence of events.
  • Reification: A nominalization operator turns predicates into entities (e.g., the state of "being true") so the system can reason about the existence and causes of conditions.

The system uses the CG5 theorem prover to derive inferences from these axioms.

Key Insights and Ideas

The "Minimalist" Metaphor

By defining a "space" as simply two independent scales, the authors show how the same logic can handle physical 3D space and metaphorical spaces (like a "space of possibilities").

Granularity and Indistinguishability

The authors introduce the idea of "grains" of knowledge. Something might be "indistinguishable" at one level of detail but distinct at another. This is crucial for defining "wear," which involves the loss of "imperceptible bits" of material that eventually lead to a perceptible change in shape.

Figure 1: A mapping between different levels of granularity, illustrating how distinct entities in one domain may be indistinguishable in another.

Axiomatizing "Wear"

One of the paper's highlights is the decomposition of the word "wear" into a chain of physical events: an impinging bit of material hits an object, causing a point-like particle to detach from the surface, which is caused by a force, occurring repeatedly over time.

Embracing "Black Holes"

The authors identify concepts like "relevant" and "normative" as philosophical "black holes." Rather than getting stuck trying to define them perfectly, they treat them as minimally analyzed primitives that are essential for making the rest of the system work.

Results and Discussion

The paper provides a qualitative proof-of-concept for defining complex, domain-specific words using a shared foundational logic:

  • Successful Definitions: They successfully axiomatized diverse words such as "range," "immediately," "brittle," "operate," "clog," and "wear."
  • Inference Capabilities: The system could prove simple theorems, such as how a change in a part's shape (due to wear) can lead to the failure of the device's function.
  • Selectional Constraints: Linguistic constraints (e.g., you can "clog" a pipe with sand but not a hole with a rope) emerge naturally as a by-product of the logical requirements of the predicates.

Conclusion and Future Work

While the system could prove simple theorems, the authors acknowledge that proving complex, multi-step causal chains remains a significant challenge. Additionally, the account of "relevance" and "normative behavior" is still provisional. The next step identified is the full integration of this knowledge base into real-time text processing for larger volumes of technical documentation.

Summary updated

BibTeX
@inproceedings{hobbs1986commonsense,
  author = {Jerry R. Hobbs and William Croft and Todd R. Davies and Douglas Edwards and Kenneth I. Laws},
  title = {{Commonsense Metaphysics and Lexical Semantics}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1986},
  doi = {10.3115/981131.981163},
}
Figure