Some Basic Mechanisms for Common Sense Reasoning About Stories Environments

Giovanni Adorni, Mauro Di Manzo, Fausto Giunchiglia

Abstract

This paper presents a foundational framework for common-sense reasoning within narrative environments, specifically addressing the "object instantiation" problem. Developed at the University of Genoa in the early 1980s, the NAUSICA system utilizes a set of local reasoning rules to infer the existence and spatial placement of implicit objects required to make a narrative scene physically plausible. By categorizing relationships such as part-whole dependencies, the containment of fluids, and state-maintenance conditions, the system constructs a conceptual mental model of a scene that extends significantly beyond the explicit text provided in a story.

Introduction

When humans read a narrative, they perform a massive amount of "under-the-hood" cognitive processing. If a story mentions a "branch on the roof," a reader instantly imagines a tree, a house, and likely a yard, even if those elements are never explicitly mentioned. This ability to visualize a coherent environment is a cornerstone of human intelligence.

In the early 1980s, traditional natural language processing (NLP) could parse the grammar of a sentence but lacked the "visual knowledge" necessary to understand the physical world described. This paper explores the logic required for a machine to bridge that gap, moving beyond simple text parsing to propose a system that uses common-sense rules to infer the existence and placement of objects that are never explicitly mentioned in a narrative.

The Problem of Object Instantiation

The authors tackle the challenge of object instantiation: how can a system decide which implicit objects are necessary to make a scene physically plausible, and where should they be placed in a 3D-like conceptual space? Without this capability, an AI's understanding of a story remains a sparse collection of disconnected nouns rather than a dense, navigable environment.

Methodology: The NAUSICA System

The authors describe the logic behind NAUSICA, a conversational system that uses Italian as its primary interface. The system relies on a "conceptual dependency" formalism—inspired by the work of Roger Schank—to represent knowledge. The core of the approach is a set of "local reasoning" rules categorized into several distinct mechanisms:

  • PART-OF Instantiation: If a part is mentioned (e.g., a "lock"), the system automatically infers the existence of the whole (a "door").
  • CONTAINED Instantiation: This mechanism is specifically designed for "fluids." Interestingly, the authors define fluids as both liquids and granular solids (like sand) because they share the functional property of requiring a container. The system infers a likely container based on the environment (e.g., a glass on a table versus a bucket on a street).
  • STATE CONDITIONS: These are inferences based on the survival or functional requirements of an object. For example, a "fish on a table" implies the presence of water to keep the fish in a "good physical state."
  • SUPPORT & SUB-PART: These rules handle physical stability and functional components. A picture on a wall implies a nail for support, while a pencil "in" a table implies the existence of a drawer.

To ensure these inferences do not lead to physical absurdities, the system employs Consistency Checks, such as POSITION and DIMENSION checks, to verify that inferred objects do not violate the laws of physics or common sense.

Key Insights and Principles

Several unique conceptual ideas emerge from the NAUSICA framework:

  1. The Principle of Minimum Objects: To avoid cluttering the "imagined" scene with redundant entities, the system follows a rule to avoid instantiating new objects if an existing object in the environment can satisfy the required spatial relationship.
  2. Environment-Driven Inference: Common sense is shown to be highly contextual. Water on a table triggers a "glass" inference, while water on a street does not. The system's "imagination" changes dynamically based on the setting.
  3. Implicit vs. Explicit Knowledge: The paper highlights that the majority of a scene is typically not described in a story. Reasoning serves as the bridge that allows an AI to move from a sparse sentence to a dense, logically consistent environment.

Case Study: "A Plant in the Living Room"

The paper demonstrates the system's logic through qualitative examples. For the input sentence "A plant in the living room," the NAUSICA system executes the following chain:

  1. It instantiates a room with default dimensions, doors, and windows (using the CONTENT mechanism).
  2. It populates the room with typical furniture, such as sofas and armchairs.
  3. It infers that the plant requires "earth" to maintain its physical state (STATE mechanism).
  4. It infers that the earth, being a "fluid," requires a "pot" (CONTAINED mechanism).
  5. It places the pot on the floor by default.

If a subsequent sentence states, "The plant is on the table," the system is sophisticated enough to move the existing pot to the table rather than creating a second plant, maintaining the internal consistency of the scene.

Limitations and Conclusion

While pioneering, the research acknowledges several limitations. The rules discussed are primarily restricted to static environments; the authors note that handling movement, temporal changes, and the goals of actors represents a significantly larger challenge. Furthermore, the system focuses on local reasoning—looking at individual relationships—rather than "global reasoning," where multiple conflicting constraints might exist across an entire narrative.

At the time of publication, the system was still under implementation. However, the work remains a significant early attempt to codify the "unwritten" rules of the world that allow humans to understand stories, providing a blueprint for how machines might one day possess common sense.

Summary updated

BibTeX
@inproceedings{adorni1983some,
  author = {Giovanni Adorni and Mauro Di Manzo and Fausto Giunchiglia},
  title = {{Some Basic Mechanisms for Common Sense Reasoning About Stories Environments}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1983},
}
Figure