SAPHIR + RESEDA: A New Approach to Intelligent Data Base Access

TL;DR

This paper introduces SAPHIR+RESEDA, a novel natural language interface (NLI) for databases that goes beyond simple keyword matching. By integrating a knowledge base and an inference engine, it can semantically transform user queries into more database-friendly versions, enabling "helpful understanding" and transportability across different database schemas.

What Problem Does This Paper Address?

Existing natural language interfaces (NLIs) for databases often struggle with the disconnect between how users naturally express queries and the rigid structure of database schemas. Users are frequently forced to learn complex query languages or understand the database's internal organization to get meaningful results. This paper addresses the challenge of creating NLIs that are more "cooperative" and "transportable," allowing users to ask questions in natural language without needing deep knowledge of the underlying database structure. The goal is to bridge the gap between user intent and database reality, even when direct mapping isn't possible.

Methods

SAPHIR+RESEDA employs a multi-layered approach:

  • Natural Language Processing (NLP): A parser, based on French grammar rules, analyzes the natural language query and produces a "deep predicate representation" that captures the query's meaning.
  • Knowledge Base (KB): A hierarchical KB is used, comprising:
  • KB1 (Most General): Contains universal common-sense knowledge, applicable across domains.
  • KB2 (Domain Specific): Holds knowledge for particular domains (e.g., staff management).
  • KB3 (Application Specific): Contains knowledge about the specific database schema.
  • Inference Engine (RESEDA): This is the core of the system's intelligence. It uses a classical expert system approach to:
  • Check Query Meaningfulness: Determines if the parsed query is relevant to the database schema.
  • Query Transformation: If a direct mapping isn't possible, the inference engine uses the KB to transform the query into a "semantically close" equivalent. This means the answer to the transformed query implies the answer to the original one. This transformation relies on two types of rules:
  • Transformation Rules: Convert one predicate representation into another, often using common-sense knowledge.
  • Standardization Rules: Translate parser output into a canonical form usable by transformation rules.
  • Transportability: The system aims for transportability by separating general knowledge (KB1) from domain-specific (KB2) and application-specific (KB3) knowledge.

Datasets

Not discussed in the paper. The paper focuses on the system's architecture and methodology, using illustrative examples rather than specific datasets for training or evaluation.

Key Insights and Ideas

  • "Semantically Close" Queries: The paper introduces a crucial concept: transforming a query into one where the answer implies the original answer. This allows for helpful suggestions when direct answers aren't available.
  • Transportability through Layered Knowledge: By separating general, domain-specific, and database-specific knowledge, the system can be more easily adapted to new databases and domains.
  • Cooperative Understanding: The system doesn't just reject invalid queries; it actively tries to understand user intent and suggest alternative, meaningful queries based on its knowledge.
  • Inference for Bridging Gaps: The inference engine is key to overcoming the mismatch between natural language and database schemas, using common-sense reasoning to infer relationships.

Key Results

The paper presents a conceptual framework and a prototype system. While quantitative results are not detailed, the key qualitative finding is the system's ability to handle a typical problem: asking "who has been in the US" when the database only knows birthplaces and degrees. The system can then propose looking for people born in the US or holding American degrees, demonstrating its "helpful understanding."

Limitations and Future Work

  • Prototype Stage: The system is described as a prototype under development (as of December 1984), with an "exhaustive" inference engine that might be inefficient for complex scenarios.
  • Rule Specificity: The rules are noted as being "temporarily very specific" due to limited testing on a single database.
  • Future Work: The authors plan to move towards more sophisticated inference approaches, such as choosing among multiple representations for the same question, after the prototype is thoroughly tested. They also aim to develop more general KB2s for various real-world domains.

Summary updated

BibTeX
@inproceedings{1985saphir,
  title = {{SAPHIR + RESEDA: A New Approach to Intelligent Data Base Access}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1985},
}
Figure