Artificial Intelligence Corporation

J. Norwood CroutACL 1981 · DOI

Author: J. Norwood Crout, Artificial Intelligence Corporation

Abstract

This paper provides a technical overview of INTELLECT, one of the first commercially viable natural language interface systems for databases. Developed in the late 1970s, INTELLECT bridges the gap between non-technical users and complex data structures by translating plain English into database commands. We describe the system's architecture, including its custom lexicon and translation engine, and evaluate its performance using both synthetic and real-world datasets. Our findings highlight the importance of handling "terse" or ungrammatical inputs and the system's success in enabling data retrieval for casual users.

Introduction

In the early era of database management systems (DBMS), accessing data required significant technical expertise. Users were typically required to master structured query languages and understand underlying data models, creating a bottleneck for business decision-makers. The INTELLECT system addresses this "man-machine communication problem" by allowing users to query databases using "true English." The goal was to create a system capable of understanding the variety of ways a human might phrase a request—including fragments and "un-English" shorthand—and translating those requests into precise retrieval commands for systems like ADABAS or VSAM.

System Architecture and Methodology

The Lexicon

The foundation of INTELLECT is the Lexicon, a custom dictionary prepared for every database implementation. The Lexicon maps specific business vocabulary to data items and describes the relationships between these words and the database schema. This mapping allows the system to resolve domain-specific terminology into actionable database queries.

Translation and Query Generation

The core logic of INTELLECT is its translation engine, which converts typed English queries into retrieval commands for the specific DBMS in use. To ensure linguistic robustness, the system was tested using a query generation program. This program takes 20 to 50 core phrases (representing record selection or data items) and generates thousands of permutations, including variations with and without prepositions, conjunctions, or articles. This ensures the system can handle diverse syntax and maintain "fluency."

Augmented Retrieval Capabilities

Beyond simple data retrieval, INTELLECT was designed to handle complex interactions:

  • Anaphora and Context: The system can resolve references to previous queries, allowing for conversational data exploration.
  • Arithmetic Operations: INTELLECT supports calculations such as totals, maximums, and other basic functions on numeric fields.
  • Formatting: The system can sort and categorize records into structured reports for better presentation.

Evaluation and Results

Datasets

The system's performance was evaluated using two primary sources:

  1. Synthetic QA Files: A file of queries generated via permutations of core phrases to test linguistic robustness.
  2. Real-world Logs: Analysis of several hundred queries collected from two active customer sites to measure successful processing rates in actual business environments.

Key Results

Quantitative analysis of customer logs demonstrated a high rate of successful processing, indicating that users could communicate effectively with the system. Qualitative feedback showed high user enthusiasm for the system's ability to provide data access without formal training.

A significant finding was the "Terse Query" phenomenon: experienced users often abandon "proper" English in favor of fragments to save keystrokes. This insight led to the realization that a successful system must prioritize understanding these "terse" inputs over enforcing strict grammatical rules.

Discussion and Limitations

While INTELLECT offers significant advantages, it is subject to certain constraints. The system's understanding is strictly limited by the "retrieval semantics" of the underlying DBMS and the pre-defined Lexicon. Furthermore, a "conceptual gap" often exists where naive users struggle to distinguish between the data itself and the system's functional capabilities, leading to requests that fall outside the system's domain.

Conclusion

INTELLECT represents a pioneering step in natural language database interfaces. By focusing on both linguistic fluency and functional "augmented" capabilities like math and sorting, the system provides a practical solution for business data access. Future work aims to expand the system's semantic domain and identify the specific business environments best served by natural language retrieval.

Summary updated

BibTeX
@inproceedings{crout1981artificial,
  author = {J. Norwood Crout},
  title = {{Artificial Intelligence Corporation}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1981},
  doi = {10.3115/981923.981932},
}
Figure