Inference in a Topically Organized Semantic Net

Johannes de Haan, Lenhart K. SchubertAAAI 1986

Introduction

In the mid-1980s, artificial intelligence research encountered a significant scaling hurdle known as the "symbol-mapping problem." While logic-based systems excelled at formal reasoning, they often faltered when tasked with "obvious" common-sense inferences. The sheer volume of facts in a large knowledge base created a "needle-in-a-haystack" scenario, where theorem provers were overwhelmed by irrelevant data, leading to a combinatorial explosion of logical steps. To address this, de Haan and Schubert proposed ECOSYSTEM, a semantic network that utilizes topical organization to streamline deductive inference.

Architectural Pillars

The ECOSYSTEM architecture is built upon the premise that indexing is a form of inference control. By restricting the search space to relevant topics, the system avoids the vast majority of unnecessary logical substitutions.

1. Topical Organization and Access Skeletons

Knowledge is partitioned into a three-level hierarchy: Subnets (specific "worlds" or stories), Concepts (entities like "Wolf"), and Topics (sub-classifications like "feeding" or "appearance"). To optimize memory and speed, the system employs "Access Skeletons." These are minimal versions of the global hierarchy that only include branches relevant to the current subnet.

The Topic Access Skeleton (TAS) provides a structured hierarchy for categorizing facts, from physical qualities to social behaviors.

2. Generalized Resolution

Rather than relying solely on traditional step-by-step logic, ECOSYSTEM uses "generalized resolution." This allows the system to take shortcuts in common domains. For instance, it can immediately resolve a conflict between "Wolf1 is Brown" and "Wolf1 is Grey" because it possesses built-in knowledge that these colors are mutually exclusive, bypassing the need to derive this from first principles.

3. Agenda-Based Control

The inference engine manages an "agenda" of potential resolutions, prioritized by their estimated computational cost. This ensures that the simplest and most direct logical paths are explored first, mimicking the rapid retrieval speeds observed in human cognition.

The Little Red Riding Hood Knowledge Base

The system was evaluated using a custom dataset based on the story of Little Red Riding Hood (LRRH). This knowledge base contained over 200 logical clauses defining the characters, objects, and physical rules of the tale.

A pruned Concept Access Skeleton (CAS) for the Little Red Riding Hood subnet, illustrating how the system isolates relevant entities.

The test set consisted of 40 questions of varying complexity, ranging from basic type checks ("Is there a creature?") to more intricate physical constraints ("Does the Wolf live in the basket?"). By using the Concept Hierarchy, the system could prune invalid resolutions—for example, it would never attempt to substitute a "GIRL" into a variable restricted to the "WOLF" branch.

Results and Discussion

The experimental results demonstrated the efficiency of topically organized inference. The system processed the 40-question test set in approximately 15 seconds on a VAX 11/780. Most significantly, the authors observed "scale invariance": doubling the size of the knowledge base had no discernible impact on question-answering time. This confirms that topical indexing effectively shields the inference engine from the overall volume of the database.

While successful, the current implementation is restricted to first-order predicate calculus and relies on hand-coded "major implications" for physical reasoning. Future work aims to incorporate temporal logic, arithmetic, and a natural language interface to further bridge the gap between formal logic and human-like understanding.

Summary updated

BibTeX
@inproceedings{haan1986inference,
  author = {Johannes de Haan and Lenhart K. Schubert},
  title = {{Inference in a Topically Organized Semantic Net}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}

Loading related content...

Figure