Two Semantic Worlds: A Data Base System with Provision for Natural Language Input

Piero Dell'Orco, Margaret King, V. N. Spadavecchia

In the mid-1970s, the emergence of relational databases, pioneered by E.F. Codd, promised a more structured way to manage information. However, a significant barrier remained for non-technical users: the complexity of formal query languages. Early attempts at Natural Language Interfaces (NLIs) were notoriously "brittle"—tightly coupled to specific database schemas. Any change to a table name or the addition of a new data category necessitated a complete overhaul of the language processor.

In "Two Semantic Worlds," the authors propose a modular solution to this portability problem. Their core thesis is the necessity of distinguishing between two distinct "semantic worlds": the general linguistic world (how language functions) and the specific database world (the entities and relations within a particular dataset).

The "Two Worlds" Philosophy

The authors argue that a robust NLI must separate general world knowledge from the restricted domain of a database. This distinction acknowledges that while a word might have a clear meaning in English, its "meaning" within a database is strictly defined by the available data and schema. By decoupling these layers, the authors aimed to create a system where the underlying data could change without requiring a total redesign of the natural language processor.

System Architecture and Methodology

The prototype system is built upon a relational model and utilizes a multi-layered translation process to convert English queries into formal commands. The architecture is divided into two primary modules:

  1. The Relational Foundation: Data is organized using Codd’s relational model. Users can interact with the system via a formal query language called AQL (interpreted into APL procedures), which serves as the target for the natural language translation.
  2. Module 1: General NL Analyzer: This module handles the "first world" of semantics. It utilizes a version of Yorick Wilks’ Preference Semantics to generate an intermediate semantic representation of the user's input. Crucially, this module is independent of the database schema; it focuses on linguistic structure and general meaning.
  3. Module 2: Database Semantic Component: This module addresses the "second world." It maps the intermediate representation from Module 1 onto the specific relations and domains of the target database.
  4. The Transformation Algorithm: Finally, the system recodes the mapped semantic representation into a formal AQL query.

This separation ensures that if a database schema is modified, only the mapping in Module 2 requires updating. The complex linguistic analyzer in Module 1 remains untouched, significantly improving system portability.

Prototype and Implementation

To demonstrate the architecture's flexibility, the authors utilized two distinct sample databases for their prototype:

  • Energetic Resources: A database focused on energy statistics.
  • Department Store: A model containing information on items, suppliers, employees, and departments.

The implementation of the formal query language (AQL) and the core database management system was completed at the time of publication. The authors successfully validated the design of the NL analyzer and the encoding algorithms, ensuring they were unaffected by changes in the underlying database structure.

Key Insights and Limitations

A key insight of the research is how the database schema itself can assist in semantic disambiguation. For instance, the verb "to sell" can help the system determine that the word "shoe" refers to an "item" domain rather than another potential meaning. This allows the database context to resolve linguistic ambiguities that the general analyzer might struggle with.

However, the system was presented as a work in progress. While the AQL and database components were functional, the Natural Language analyzer was still under development and not yet fully integrated at the time of writing. Residual ambiguities in reference and inference remained, and the system's portability was still constrained by the initial vocabulary of the general linguistic module; if a user uses words outside that set, the first module requires manual enlargement.

Conclusion

By explicitly splitting general linguistic knowledge from database-specific semantics, the authors moved toward a "plug-and-play" architecture for natural language interfaces. This modular approach addressed the brittleness of early systems and laid the groundwork for more portable and user-friendly database interactions for non-specialists.

Summary updated

BibTeX
@inproceedings{dellorco1977two,
  author = {Piero Dell'Orco and Margaret King and V. N. Spadavecchia},
  title = {{Two Semantic Worlds: A Data Base System with Provision for Natural Language Input}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1977},
}
Figure