A Partial Taxonomy of Knowledge about Actions

Eugene CharniakIJCAI 1975

Introduction

In his foundational 1975 paper, Eugene Charniak addresses a critical challenge in artificial intelligence: the representation of common-sense knowledge. For a machine to truly understand human stories, it requires more than a dictionary; it needs a structured understanding of the logic behind human actions. Charniak proposes a "partial taxonomy" that categorizes this knowledge along two primary dimensions: Force (the motivation) and Form (the structure).

The Challenge of Common Sense

Human readers effortlessly infer the logic behind social interactions. For instance, if a character decides not to buy a gift because another person already has it, we understand the underlying social convention. A computer, however, lacks this "how-to-do-it" knowledge. Charniak’s goal was to organize this vast, messy web of activity into a framework that allows language comprehension programs to make useful inferences about character motivations and physical constraints.

The Two-Dimensional Taxonomy

Charniak’s framework is built on two axes:

1. Force: The "Why"

Force describes the motivational power or "driving power" behind a rule. Charniak identifies four distinct types:

  • Strict: Logical or physical necessities. For example, you must have an apple to eat it.
  • Social: Cultural norms and etiquette, such as announcing your name on the phone.
  • Suggested: Practical advice for efficiency or safety, like putting paper on the floor before painting.
  • Regulatory: Rules enforced by authority. Breaking these doesn't physically stop the action but "invalidates" the result (e.g., an illegal forward pass in football).

2. Form: The "What"

Form describes the structural nature of the information. This includes Subactions (the steps in a process), Substates (conditions that must hold during an action), and Time Orderings (the sequence of events).

Case Study: The Umbrella Method

To demonstrate the taxonomy, Charniak uses the act of using an umbrella as a primary case study. This example highlights how physical components interact with logical states and different types of forces.

Anatomy of an umbrella (D1), providing the structural basis for the action-state analysis.

As shown in the diagram above, the umbrella consists of several parts—the canopy, stick, and catches—each playing a role in the "Umbrella Method." The taxonomy allows us to define the "Strict" force required to raise the canopy (pushing the flange) versus the "Suggested" force of pointing the stick away from one's eyes for safety.

Integrating Force and Form

A key insight of the paper is that almost any Force can combine with any Form. We can observe social subactions (hosting a party after a wedding) or strict subactions (mechanical steps to open a device).

Logical flow for opening an umbrella (D3-D5) and examples of social and regulatory rules (F27-F35), illustrating the breadth of the taxonomy.

The figure above illustrates the complex interplay of these categories. It maps out the substates (e.g., the canopy must be between the rain and the object) and the methods for achieving them. It also lists various social and regulatory rules, demonstrating the framework's versatility across physical and social domains.

Reasoning About Broken Rules

Charniak offers an elegant heuristic for how humans reason about deviations from these rules. If a Strict prerequisite is broken, we ask "How?" (e.g., "How did he fly a kite without wind?"). If a Non-strict (Social or Suggested) rule is broken, we ask "Why?" (e.g., "Why did he paint the chair without putting paper down?"). This distinction is crucial for AI systems attempting to model human curiosity and narrative comprehension.

Conclusion

While Charniak acknowledges his taxonomy is "partial" and that boundaries between categories can be fuzzy, the work provides a vital roadmap. By distinguishing between the physical "how" and the motivational "why," Charniak laid the groundwork for more sophisticated common-sense reasoning engines in AI, moving beyond simple fact-retrieval toward true narrative understanding.

Summary updated

BibTeX
@inproceedings{charniak1975partial,
  author = {Eugene Charniak},
  title = {{A Partial Taxonomy of Knowledge about Actions}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1975},
}
Figure