A Knowledge Framework for Natural Language Analysis

Paul S. Jacobs

Introduction

In the late 1980s, Natural Language Processing (NLP) faced a significant hurdle known as the Role Relationship Problem. While identifying the subject and object of a sentence was straightforward, determining the underlying semantic meaning of those roles in context remained difficult. For example, the verb "send" implies a human recipient in "Jones sent the message to Frank," but a physical destination in "The job was sent to the line printer." Metaphorical usage, such as "The command takes two arguments," further complicates this, as a computer command does not "take" objects in a literal, physical sense.

Traditional systems often required hard-coding every specific sense for every verb. Paul S. Jacobs proposed the Ace framework to address this inefficiency by capturing the underlying patterns of human language through hierarchical associations and metaphorical mappings.

The Ace Framework

The Ace framework organizes linguistic and conceptual knowledge into a structured hierarchy using three primary components:

  1. Linguistic Relations: A hierarchy of syntactic structures that recognizes general behaviors (e.g., how an "indirect object" functions) across different verbs.
  2. REF Associations: These links connect linguistic roles to conceptual roles. Through inheritance, a system can apply general knowledge about a "recipient" to specific instances like a "customer" in a sales event.
  3. VIEW Associations: This mechanism handles metaphors by mapping one conceptual domain onto another. For instance, a "command execution" can be viewed as a "transfer," allowing the system to process "taking an argument" using the same logic as "taking a gift."
VIEW relationship mapping command execution to a transfer event

The TRUMP Analyzer

The Ace framework is implemented in the TRUMP analyzer (TRansportable Understanding Mechanism Package). TRUMP processes natural language in three distinct phases:

  • Syntactic Matching (Parsing): Identifying the syntactic structure of the input.
  • Mapping: Utilizing REF and VIEW links to determine conceptual roles.
  • Concretion: Narrowing a general concept (e.g., "transfer") to a specific domain-relevant one (e.g., "sending electronic mail") based on context.
The three-phase process of natural language analysis in TRUMP

Key Insights

  • Metaphor as a Systematic Rule: Rather than treating metaphors as edge cases, Ace uses VIEW associations to capture the consistent logic behind metaphorical expressions.
  • Knowledge Inheritance: Organizing roles into a hierarchy reduces manual data entry. If a "customer" is defined as a type of "recipient," the system automatically inherits the relevant properties.
  • Portability: By establishing a "core" of linguistic knowledge, the TRUMP analyzer was designed to be moved between disparate domains—such as office automation and military command—without requiring a complete rewrite of the knowledge base.

Results and Limitations

The TRUMP analyzer was implemented in approximately 5,000 lines of Common Lisp, with a core knowledge base of several hundred linguistic entries and a thousand concepts. Qualitatively, the system successfully distinguished between different senses of verbs like "send" and "give" using selectional restrictions.

However, a "knowledge bottleneck" remains; moving to a new domain still requires significant manual knowledge engineering. Additionally, the complexity of VIEW associations makes full conceptual networks difficult to visualize and manage. Future work is needed to improve ambiguity resolution in cases where multiple interpretations are equally plausible.

Summary updated

BibTeX
@inproceedings{jacobs1987knowledge,
  author = {Paul S. Jacobs},
  title = {{A Knowledge Framework for Natural Language Analysis}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1987},
}
Figure