Natural Semantics in Artificial Intelligence

James R. Carbonell, Allan M. Collins

Introduction

In the early 1970s, the field of Artificial Intelligence was largely dominated by systems operating within "closed worlds"—narrow, highly structured domains like airline reservations or geometric block-stacking. In these environments, every fact was either explicitly known or definitively false. However, human cognition operates in an "open world," characterized by imprecision, incompleteness, and uncertainty. This foundational 1973 research argues that for AI to achieve true natural language understanding, it must move beyond rigid, two-valued logic and embrace the "fuzzy" nature of human knowledge.

The Problem: Closed Worlds vs. Open Realities

The authors identify a significant gap between the rigorous logic of computers and the flexible, often ill-defined semantics of human thought. Humans effectively navigate information that is:

  • Imprecise: Understanding that a "large salary" is high without requiring a specific dollar amount.
  • Incomplete: Recognizing that we do not know everything, yet possessing the meta-knowledge to understand what we should know.
  • Uncertain: Asserting truths in relative terms rather than absolute booleans.

To bridge this gap, the researchers explored how a machine might handle questions it wasn't explicitly programmed to answer, or how it might admit ignorance based on the depth of its own expertise.

Methodology: The SCHOLAR System

The primary vehicle for this research was SCHOLAR, a "mixed-initiative" instructional system designed for tutoring. Unlike its contemporaries, SCHOLAR did not rely on a simple database of facts. Instead, it utilized several technical innovations to simulate human-like reasoning:

Extended Semantic Networks

Knowledge in SCHOLAR is stored as a network of facts and concepts, enhanced with I-tags (Importance tags). These tags allow the system to determine the saliency of information within a given context. For instance, when discussing a country, the system uses I-tags to prioritize major geographical features over minor villages, mimicking human conversational priorities.

Open-World Inference Strategies

To handle the gaps in its knowledge base, SCHOLAR employs a "bag of inferential tricks" modeled after human protocols:

  • Negative Inferences: Rather than assuming "not found = false," the system uses a "lack-of-knowledge" strategy. If the system possesses extensive knowledge about a region (e.g., Uruguay) but finds no mention of a specific product (e.g., oil), it infers that the product is likely not produced there.
  • Functional Inferences: The system performs "quasi-calculations," such as inferring a climate based on a location's latitude and altitude.
  • Functional Analogies: SCHOLAR can reason by comparison—for example, suggesting that Region A might have cattle because it is topographically and climatically similar to Region B, which is known for cattle ranching.

Datasets and Evaluation

The methods were evaluated using two primary knowledge bases:

  1. Geography of South America: The main testbed, containing complex data on countries, cities, climates, and economic products.
  2. The ARPA Network: A secondary database focused on the technical specifications of the early internet.

To refine these strategies, the authors analyzed tape-recorded dialogues between human tutors and students, identifying the specific heuristic shortcuts humans use to fill information gaps.

Key Insights: Fuzziness as a Feature

The paper posits that "impreciseness" is not a bug in human language but a feature that makes communication more efficient. By modeling this fuzziness, AI can become more robust. A central discovery was the importance of Meta-Knowledge: humans reason about their own knowledge. We know that if a fact were important, we would likely know it; therefore, if we do not know it, we can reasonably infer it is either false or insignificant.

Furthermore, the authors found that certainty is rarely the result of a single logical chain. Instead, it arises from Combining Inferences. If a deductive path and an analogical path both lead to the same conclusion, the system’s confidence in that answer increases, mirroring the human experience of "triangulating" truth.

Results and Qualitative Performance

Through sample protocols, the authors demonstrated SCHOLAR’s ability to provide nuanced, human-like interactions:

  • Contextual Adaptation: The system can pivot between topics (e.g., moving from Argentina to Peru) while maintaining the thematic thread of the conversation.
  • Error Correction: By checking the semantic network's structure, the system identifies category errors, such as a student mistakenly identifying a religion as a country.
  • Handling Fuzzy Queries: When asked if a city is "large," SCHOLAR does not look for a binary attribute. It compares the city’s population to others in the same region to provide a relative, context-aware answer.

Limitations and Future Outlook

While groundbreaking, the authors noted several limitations. SCHOLAR’s English syntax parsing remained primitive, and the system lacked true inductive reasoning—the ability to learn new rules from examples. Additionally, while U-tags (Uncertainty tags) were proposed to represent data reliability, they were not fully implemented at the time. The authors estimated that humans use approximately one hundred different inferential strategies, suggesting that scaling this "bag of tricks" approach would require significant future effort.

Ultimately, the paper remains a landmark in AI for its insistence that true intelligence requires the ability to reason effectively with incomplete and imprecise information.

Summary updated

BibTeX
@inproceedings{carbonell1973natural,
  author = {James R. Carbonell and Allan M. Collins},
  title = {{Natural Semantics in Artificial Intelligence}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1973},
}
Figure