Knowledge Structures and Language Boundaries

Yorick Wilks

Abstract

Presented at the Natural Language-8 conference in 1977, Yorick Wilks’ work marks a pivotal shift from rigid, rule-based linguistics to flexible, knowledge-driven AI. This paper introduces "Preference Semantics" and "pseudo-texts" to address the inherent flexibility of human language, specifically how machines can interpret metaphorical or "extended" usage—such as "My car drinks gasoline"—without failing due to strict semantic violations.

Introduction: Beyond Selection Restrictions

In the 1970s, Natural Language Understanding Systems (NLUS) typically relied on rigid semantic rules. For instance, a dictionary entry for the verb "drink" might strictly require an animate subject. Encountering a sentence like "My car drinks gasoline" would cause such systems to flag a semantic error and halt.

Wilks argues that this approach fundamentally misinterprets human communication. Metaphor, metonymy, and extended use are not exceptions but the norm. The core challenge addressed is how a machine can gracefully handle these "preference violations" to reach a correct interpretation by projecting world knowledge onto the text.

Methodology: Preference Semantics and Pseudo-texts

Wilks’ system relies on several innovative components to achieve this flexibility:

  1. Semantic Primitives: Actions and entities are decomposed into basic building blocks. For example, the verb "drink" is represented as a structure involving a cause to move a liquid substance into a self-entity.
Semantic primitive structure for the verb "drink," illustrating the CAUSE to MOVE FLOW STUFF into a SELF.
  1. Pseudo-texts (PTs): These are frame-like knowledge structures. Uniquely, Wilks designed PTs to share the exact same format as the system’s internal representation of input text. This homogeneity allows the same inference engine to process both background knowledge and new inputs seamlessly.
  2. Sense-Projection: When a preference is violated (e.g., an inanimate "car" drinking), the system consults the PT for the subject. Finding that cars use liquid fuel to move, it "projects" this knowledge onto the verb, effectively re-interpreting "drink" in that specific context as "consume fuel."
  3. Thesaurus-Based Hierarchy: The system employs a hierarchical thesaurus to organize formulas, allowing it to attempt matches at specific levels (e.g., "ambulance") and generalize upward (e.g., "vehicle") if necessary.

Key Insights: Language as the Model for Knowledge

A central tenet of Wilks’ philosophy is that metaphor is the standard state of language. Consequently, AI must be built to expect and resolve violations rather than reject them. Furthermore, Wilks argues for homogeneous representation: since representing language is the most difficult task in AI, world knowledge should be assimilated into text-like structures rather than forcing language into abstract logic. Finally, he advocates for a "Lazy System" assumption, where deep inference is only triggered when an ambiguity or preference violation occurs.

Results and Disambiguation

The system's effectiveness is demonstrated through qualitative analysis of complex linguistic "edge cases," such as coreference resolution in narratives involving accidents.

A pseudo-text (PT) representation showing the interconnected knowledge used to resolve complex linguistic scenarios like a car "using" liquid.

For example, in the sentence "An ambulance... collided with a car killing its driver," the system correctly identifies "its" as referring to the car. It achieves this by checking "traffic-lights" pseudo-texts and applying pragmatic overrides to avoid redundant or illogical interpretations. It also distinguishes between "plausible" violations (a car drinking) and "implausible" ones (a car carving a roast) by checking for overlapping semantic structures in its PTs.

Limitations and Future Directions

While groundbreaking, the system faced significant hurdles:

  • Computational Complexity: Matching complex PTs remains a difficult problem, especially with negation or abstract concepts.
  • Scalability: The reliance on hand-coded formulas and PTs made scaling to the entirety of the English language a massive undertaking.
  • Idioms: Certain idiomatic expressions may require specialized dictionary entries rather than general inference.

Wilks suggested that future work should move toward question-answering environments to better evaluate the relative usefulness of different knowledge branches.

Summary updated

BibTeX
@inproceedings{wilks1977knowledge,
  author = {Yorick Wilks},
  title = {{Knowledge Structures and Language Boundaries}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1977},
}
Figure