A Hybrid Approach to Representation in the Janus Natural Language Processor

Ralph M. WeischedelACL 1989 · DOI

Abstract

Janus is a pioneering natural language processing (NLP) system designed to bridge the gap between expressive intensional logic and efficient taxonomic reasoning. Developed by Ralph M. Weischedel and colleagues at BBN Systems and Technologies, Janus utilizes a hybrid representation to model human language. By delegating complex semantics to a World Model Language (WML) and domain-specific ontology to a taxonomic language (NIKL), the system achieves high performance without sacrificing depth. Evaluation across a 7,000-word lexicon reveals that 99.9% of domain vocabulary can be successfully modeled within the taxonomic framework, proving the viability of this "best of both worlds" approach.

Introduction: The Expressivity-Efficiency Trade-off

In the late 1980s, NLP research faced a fundamental trade-off. Highly expressive frameworks like First-Order Logic (FOL) or Intensional Logic could handle complex nuances—such as time-varying facts, hypothetical "what-if" scenarios, and quantifiers like "most" or "always"—but were computationally prohibitive for real-time interaction. Conversely, taxonomic languages (frame-based systems) offered efficient inheritance reasoning and hierarchical organization (e.g., "a Destroyer is a type of Ship") but struggled with linguistic subtleties, such as distinguishing between specific sets ("the cats") and general concepts ("cats"). Janus addresses this by using a hybrid architecture that assigns different aspects of language understanding to the most suitable tool.

The Janus Architecture

The Janus system relies on two primary components linked by a mapping layer:

  1. World Model Language (WML): An intensional logic used for the semantic representation of utterances. It manages time indices, possible worlds, and propositional attitudes (e.g., "I believe that...").
  2. NIKL (New Implementation of KL-ONE): A taxonomic representation language used to define the domain ontology, organizing concepts (unary predicates) and roles (binary relations).
Taxonomic representation of the relationship "CATS EAT MICE" in Janus.

To maintain speed, Janus avoids complex theorem proving. Instead, it pre-compiles inferential chains within the NIKL taxonomy. For example, determining if a "USS Enterprise" can be the object of a "command" involves a simple check of the taxonomic hierarchy to see if "Carrier" is a descendant of "Military Unit."

Key Insights: Efficiency and Scalability

A significant finding of the Janus project was the "99.9% Rule": nearly all domain-specific vocabulary in a 7,000-word lexicon could be defined using the limited expressive power of the NIKL taxonomy. Only 0.1% of words, such as "upgrade" or "expect," required the full power of intensional logic.

Efficiency is further enhanced through operator-driven modeling. By utilizing logical operators like THE, POWER, KIND, and SAMPLE, the system can derive complex entities—such as definite plurals or generic classes—from a single taxonomic constant. This approach keeps the knowledge base lean and manageable.

Mapping of logical expressions to semantic classes using operators like THE, POWER, and KIND.

Results and Future Directions

The hybrid approach demonstrated remarkable scalability and effectiveness in resolving polysemy. For instance, in the sentence "USS Frederick has a speed of 30 knots," the system uses the taxonomy to identify the valid relationship between "Ship" and "Speed." Furthermore, the NIKL classifier provides a built-in consistency check, ensuring new domain knowledge does not contradict existing axioms.

Despite its successes, challenges remain. Words involving complex temporal changes or intensional arguments still require manual annotation. Future work aims to leverage the taxonomy for handling metonymy and "bridging" inferences (e.g., inferring a "ceiling" from the mention of a "room"). While automated taxonomy building using dictionaries like LDOCE faced hurdles due to circular definitions, the Janus framework remains a robust model for hybrid knowledge representation in NLP.

Summary updated

BibTeX
@inproceedings{weischedel1989hybrid,
  author = {Ralph M. Weischedel},
  title = {{A Hybrid Approach to Representation in the Janus Natural Language Processor}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1989},
  doi = {10.3115/981623.981647},
}
Figure