A Logical-Form and Knowledge-Base Design for Natural Language Generation

Norman K. Sondheimer, Bernhard NebelAAAI 1986

Introduction

In the mid-1980s, the field of Natural Language Generation (NLG) faced a significant technical hurdle known as the "impedance mismatch." On one side were reasoning systems that operated using the cold, rigid symbols of formal logic. On the other were human users who required fluid, context-dependent natural language. Bridging this gap required more than just a dictionary; it required a system that could translate abstract logical demands into grammatically correct sentences without forcing the reasoning engine to become a linguistic expert.

This 1986 paper by Sondheimer and Nebel, presented at AAAI, introduced a foundational architecture for this translation. By utilizing a hybrid knowledge representation system as a bridge, the authors demonstrated how to map domain-specific concepts to linguistically motivated categories, enabling a general-purpose grammar generator to "understand" abstract logic efficiently.

The Problem: Logic vs. Linguistics

The dilemma for early NLG researchers was two-fold. If the input to a generator was linguistically oriented, the upstream reasoning software had to possess deep knowledge of grammar—deciding, for instance, whether a verb should be active or passive. Conversely, if the input was purely logical, the generator struggled to interpret it because standard first-order logic does not inherently contain linguistic cues like tense, aspect, or voice.

The goal was to create a reusable module capable of taking well-defined formal logic and producing colloquial English. This required a system that could handle both the "connection" (the relationships between objects) and the "scope" (quantifiers like "every" or "some") of a logical statement separately, mirroring how human language functions.

System Architecture: Nigel, NIKL, and KL-TWO

The researchers integrated three distinct technologies to solve this problem:

  1. Nigel: A large-scale English grammar generator based on Systemic Functional Linguistics. Nigel operates by traversing a "chooser" tree, making grammatical decisions based on "inquiries" about the data.
  2. NIKL (New Implementation of KL-ONE): A frame-based knowledge base. The authors' key innovation was adding a "middle layer" to this network that categorized domain-specific concepts (e.g., "Sending a Mail") into linguistic categories (e.g., "Material Action").
  3. KL-TWO: A hybrid reasoner that combined NIKL with PENNI, a fast propositional reasoner. This allowed for efficient querying during the generation process.

The Workflow of Expression

The generation process begins with a demand for expression in first-order predicate calculus. This logic is decomposed into two primary components: a set of assertions describing the participants and actions, and a scope structure handling the logic of quantifiers.

Logical form representation of an action occurrence, showing the separation of existential quantifiers and relational assertions.

As Nigel builds the sentence, it queries KL-TWO to resolve grammatical choices. For example, if the logic describes a "Send" event, KL-TWO uses the NIKL hierarchy to inform Nigel that this is a "Material Action," triggering the appropriate syntactic structures in the output.

The "Whorfian" Intermediate Layer

The most significant insight of the paper was the insertion of a "systemic" intermediate layer within the knowledge base. This layer acted as a linguistic ontology. A domain expert could define a concept like "Email" according to technical requirements, but by linking it to a "Message" category in the intermediate layer, the generator automatically inherited the knowledge of how to discuss it in English.

The NIKL hierarchy acting as a linguistic intermediate layer, mapping domain concepts to process types.

This separation allowed the system to mirror human language's distinct handling of syntax and determiners. By splitting the "connection" from the "scope," the authors ensured that the logical precision of the input was preserved while the output remained natural.

Results and Efficiency

The system was tested using a domain model from the Consul project, which managed mail and calendar systems. The model included over 850 concepts and 150 roles. Despite the complexity, the hybrid reasoning approach proved remarkably efficient. By using the restricted PENNI reasoner instead of a general-purpose theorem prover, the overhead of interpreting logic was kept to just 10-20% of the total generation time.

The system successfully generated a wide range of complex structures, from simple assertions ("Smith sent a message") to quantifier-heavy phrases ("The read statuses of the messages") and relative clauses. It even demonstrated an ability to handle follow-up questions in dialogue, maintaining context across multiple interactions.

Conclusion and Legacy

While the system had limitations—such as a struggle with linguistic ambiguity and constraints on "disconnected" logical structures—it served as a vital stepping stone. The work laid the groundwork for the Janus and Penman projects, which pushed the boundaries of higher-order logic and multi-sentence text generation. Sondheimer and Nebel's design proved that a well-structured knowledge base could serve as the essential translator between the rigid world of logic and the fluid world of human speech.

Summary updated

BibTeX
@inproceedings{sondheimer1986logical,
  author = {Norman K. Sondheimer and Bernhard Nebel},
  title = {{A Logical-Form and Knowledge-Base Design for Natural Language Generation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure