Asking a Database in Plain English: A 1989 Patent That Posed the NL2SQL Question
About research memos: This entry records a candidate at the stage of confirmed source URL. The full patent description and line-by-line Claim 1 have not been read. Only confirmed facts are stated; inferences are marked as such.
Why dig here
"Ask ChatGPT about your data and it writes the SQL for you" became a real workflow in 2023. But the question — can a non-technical user search a database using plain language? — was posed as a patent in 1989. Then, the proposed answer was a knowledge base plus expert system. Now, the answer is a large language model. The same question, 34 years apart, with structurally different solutions. That gap is worth mapping.
Basic information
- Patent number: US5197005A
- Title: Database retrieval system having a natural language interface
- Filed: May 1, 1989
- Granted: March 23, 1993
- Inventors: Steven Shwartz, Claudio Fratarcangeli, Richard E. Cullingford, Gregory S. Aimi, Donald P. Strasburger (5 inventors)
- Original Assignee: Intelligent Business Systems
- Primary source: Google Patents (URL confirmed; Abstract and core description retrieved)
- Legal status: Expired (Fee Related)
What the patent describes (from Google Patents)
Two problems are stated:
The synonym problem: A user says "customers" but the database column is called "clients." The vocabulary of natural language and the vocabulary of the database designer don't match.
The data location problem: A user says "show me sales" but the system doesn't know which table, which column, which join that requires.
The patent's solution centers on a knowledge base — a structured description of the database's schema and its semantic meaning, maintained by a developer.
From the abstract:
A database developer creates a knowledge base containing structural and semantic descriptions of the application database to be queried. A database-independent standard internal semantic representation is generated, and an expert system identifies the required database elements based on this representation. A debug feature derives an external semantic representation from the internal one, presenting it in a format the developer can easily understand.
The processing flow:
- Developer builds a knowledge base (DB schema + semantic descriptions)
- User enters a natural language query
- System generates an internal semantic representation (intermediate, DB-language independent)
- Expert system maps semantic representation to DB elements and executes query
- Debug feature lets developer verify how the query was interpreted
That fifth point — a verification interface for query interpretation — overlaps in intent with modern UX patterns where users confirm LLM-generated SQL before execution.
Connections to modern systems (hypotheses)
| US5197005A (1989) | Modern technology | Assessment (pre-full-read hypothesis) |
|---|---|---|
| Knowledge base (schema + semantic descriptions) | RAG context injection / schema-aware prompting | Similar (informing the system about DB structure in advance) |
| Internal semantic representation (DB-independent intermediate) | LLM's internal intent representation | Analogy (directionally similar idea of an intermediate representation) |
| Expert system identifies DB elements | LLM generates NL2SQL | Similar (same problem: natural language → DB operation; solution design is fundamentally different) |
| Debug feature for query interpretation verification | User-facing SQL inspection UX | Similar (transparency of interpretation is the shared intent) |
| Rule-based parsing | Probabilistic language generation | Cannot map (design philosophy is entirely different) |
The most important difference: this patent's expert system applies predefined rules and a developer-built knowledge base. A modern LLM generates SQL probabilistically, without an explicit rule base. The question — natural language querying of databases — is identical. The solution architecture has no structural continuity.
These are pre-full-read hypotheses. Claim 1 details will update the table.
What's not confirmed
- Claim 1 verbatim text
- Whether the patent resulted in a commercial product (Intelligent Business Systems business history)
- Which database systems were targeted (Oracle? dBASE? Other 1989-era systems?)
- Forward citation count (influence on later NL2SQL research)
- Background of the 5 inventors
Next action
Retrieve Claim 1 and look for evidence of commercial deployment. If a product shipped, knowing what users it reached gives the article its framing. If it didn't, "why the knowledge-base approach hit a wall" is the hook.
Reference links:
- Original patent: US5197005A on Google Patents
- AI & ML Patents #3 (research note): Philips hardware backpropagation US5517598A (1993)
- AI & ML Patents Memo #5: NASA spatiotemporal neural network US5253329A (1991)