Event Shape Diagrams

David L. WaltzAAAI 1982

Abstract

Traditional natural language understanding systems often rely on coarse-grained logical primitives that strip away the subtle distinctions between similar actions. This paper introduces Event Shape Diagrams, a representational framework that maps actions, internal states, and physical consequences onto aligned timelines. By capturing timing, rate of change, and causal dependencies, this framework allows AI systems to distinguish between nuanced verbs—such as "nibbling" versus "wolfing"—and provides a structural basis for interpreting adverbs and metaphors.

Introduction: Beyond Conceptual Dependency

In the early 1980s, the dominant paradigm for semantic representation was Schank’s "Conceptual Dependency" (CD). CD reduced a wide variety of verbs to a small set of primitives; for example, eat, drink, smoke, nibble, and gulp were all represented by the single primitive INGEST.

Waltz argues that this reductionism is insufficient for deep understanding. If a system cannot distinguish between "nibbling" and "wolfing," it loses the ability to infer the agent's internal state—such as whether they are depressed, hurried, or famished. To address this, Waltz proposes a representation that captures the "shape" of an event over time.

The Event Shape Diagram Framework

The core of this approach is the Event Shape Diagram, which aligns multiple "tracks" of information along a shared timeline. Rather than treating an event as a discrete point or a simple state change, it is viewed as a set of concurrent, interacting processes.

An Event Shape Diagram for the verb 'eat', illustrating aligned timelines for action, food quantity, and internal desire.

Key components of these diagrams include:

  1. Aligned Timelines: Tracks represent different aspects of the event, such as the duration of the physical action, the quantity of an object involved, or the agent's internal desire.
  2. Scales and Typical Values: Each track operates on a scale (e.g., intensity, probability). Diagrams include "typical value ranges" or defaults, such as the standard duration of a meal versus a snack.
  3. Causal Links: Arrows indicate how changes in one track drive changes in another. For instance, the act of eating causes the amount of food inside the agent to increase, which subsequently causes the agent's hunger to decrease.

Capturing Linguistic Nuance and Adverbs

By defining meaning as a "shape" on these scales, the framework naturally handles adverbial modification. Adverbs are treated as transformations of the base diagram:

  • "Quickly" shrinks the time interval ($t_f - t_o$).
  • "A lot" increases the peak value on a quantity scale.
  • "Nibbling" vs. "Wolfing" is represented by different slopes in the rate of consumption and different initial states of "desire."

Modeling Expectation and Surprise

One of the most powerful applications of this framework is modeling cognitive states like surprise. Surprise is represented by overlaying two diagrams: one for the expected event shape and one for the actual observed event.

Representing surprise by comparing expected vs. actual event shapes for 'John ate so much'.

As shown in Figure 2, the discrepancy between the expected quantity of food consumed and the actual quantity results in a rising "interest/arousal" track. This provides a formal mechanism for an AI to "notice" when an event deviates from the norm.

Metaphor and Perceptual Grounding

Waltz suggests that metaphors function through structural mapping. In the sentence "John ate up the compliments," the "shape" of eating (desire, ingestion, internalization) is mapped onto the domain of social interaction. This allows a system to infer that John had a high "hunger" for approval and "consumed" the praise greedily.

Furthermore, because these diagrams deal with physical quantities and rates, they offer a path toward perceptual grounding. The shapes could potentially be derived directly from sensor data (vision or touch), bridging the gap between low-level perception and high-level language.

Conclusion and Future Directions

While the framework provides a robust qualitative proof-of-concept for expressive semantic representation, several challenges remain. At the time of proposal, diagrams were manually constructed, and extending the model to abstract verbs like "believe" or "consider" is non-trivial. However, Event Shape Diagrams offer a compelling alternative to primitive-based systems, enabling common-sense inferences and a deeper structural understanding of how language describes the physical and mental world.

Summary updated

BibTeX
@inproceedings{waltz1982event,
  author = {David L. Waltz},
  title = {{Event Shape Diagrams}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1982},
}
Figure