The Use of Object-Specific Knowledge in Natural Language Processing

Mark H. BursteinACL 1979 · DOI

Abstract

This paper introduces the Object Primitive Understanding System (OPUS), a framework designed to address the "common sense" gap in natural language understanding by encoding functional knowledge about physical objects. By utilizing a set of seven "Object Primitives," the system can automatically infer missing causal links—such as recognizing that opening a bottle enables pouring its contents—without the computational burden of exhaustive logical inference.

Introduction

A significant challenge in early natural language processing (NLP) was the "mundane knowledge" gap. While humans instinctively understand that "opening a bottle" is a prerequisite for "pouring wine," late-1970s parsers struggled to connect these events. Previous attempts to solve this via massive inference generation often led to a "combinatorial explosion," where systems became bogged down in thousands of irrelevant logical deductions. This research proposes a method to make these inferences directed and grounded in the functional reality of objects.

The OPUS Framework and Object Primitives

Building upon Roger Schank’s Conceptual Dependency (CD) theory, the OPUS system introduces Object Primitives (OP) to represent the functional roles of physical items. These primitives include:

  • CONNECTOR: Objects that enable movement between regions (e.g., an open door).
  • SEPARATOR: Objects that prevent movement or sensing (e.g., a closed door).
  • SOURCE/CONSUMER: Objects that provide or "use up" other objects (e.g., a faucet or a drain).
  • RELATIONAL: Describes containment or support (e.g., a bottle).
  • SETTING/GESTALT: Handles situational contexts and complex object parts.
Figure 2: Conceptual representation of a causal chain where an action on a 'bottle' (CONNECTOR) enables subsequent transfers.

The system integrates a conceptual analyzer with "demons"—software triggers that activate when specific objects are encountered. For instance, seeing "bottle" triggers a demon that loads expectations about containment and opening into short-term memory to help parse the rest of the sentence.

Directed Inference and Causal Verification

OPUS avoids combinatorial explosion by tying inferences directly to the objects mentioned. Instead of generating all possible outcomes, the system asks specific questions: "How does a faucet relate to drinking?" This directed approach uses forward-chaining (what an action enables) and backward-chaining (preconditions) to bridge gaps in text.

Figure 6: Representation of a physical transfer (PTRANS) of wine from a container, illustrating how object-specific knowledge is encoded.

Key Results

The system demonstrates qualitative success in "causal chain construction." In the sentence "John turned on the faucet so he could drink," OPUS successfully infers that the faucet (SOURCE) provides water, which John then possesses and ingests. It also handles ambiguity resolution—identifying that a glass is filled with water at a faucet even if "water" is unmentioned—and filters "weird" sentences that violate the functional expectations of objects (e.g., pulling an elephant out of a bottle).

Conclusion

While OPUS successfully demonstrates the power of object-specific knowledge, it relies on manually hand-coded primitives, presenting a scaling challenge. Future work aims to expand the scope of these primitives to capture more abstract functions and integrate them into complex, multi-sentence narrative analysis.

Summary updated

BibTeX
@inproceedings{burstein1979use,
  author = {Mark H. Burstein},
  title = {{The Use of Object-Specific Knowledge in Natural Language Processing}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1979},
  doi = {10.3115/982163.982178},
}
Figure