Combining Production Systems and Ontologies

Martín Rezk, Werner NuttRr 2011 · DOI

TL;DR

This paper tackles the challenge of integrating two distinct knowledge representation paradigms: production systems (PS) for rule-based reasoning and description logics (DL) for ontologies. They propose a novel semantics (POS) that bridges the gap, allowing facts and rules to be expressed within an ontology language, and demonstrate its embedding into fixed-point logic for formal analysis. This work is crucial for building more robust and expressive AI systems that can reason over both dynamic knowledge and static domain models.

What Problem Does This Paper Address?

Artificial intelligence systems often need to combine different ways of representing knowledge. Production Systems (PS) are excellent for modeling dynamic processes and decision-making through rules that modify a "working memory" of facts. Think of them as the engine of an expert system. On the other hand, Description Logics (DL) are powerful for defining static domain knowledge, relationships between concepts, and performing logical reasoning over ontologies. This is like having a detailed encyclopedia of the world the system operates in.

The problem arises when we want these two systems to work together seamlessly. Traditional PS assume a single, definitive "truth" in their working memory (closed-world assumption), while DLs typically operate with an open-world assumption, meaning what isn't stated isn't necessarily false. This semantic mismatch creates challenges:

  • Rule Condition Checking: How do we reliably check if a PS rule's condition is met when the facts might be spread across a DL ontology and a working memory, and entailment (logical deduction) is involved?
  • Rule Effects (Retraction): How do we handle retracting facts (removing them) when the DL ontology might imply their existence, making simple removal problematic?
  • Rule Execution (Semantics): When should a rule fire if two different working memories can be semantically equivalent due to the ontology?

This paper aims to bridge this gap by proposing a unified framework that respects the strengths of both PS and DL.

Methods

The authors introduce a new semantics called POS (Production system and Ontologies Semantics). This semantics is designed to be:

  • Consistent: It aligns with traditional PS semantics for rule application and DL semantics for knowledge representation.
  • DL Independent: The core ideas are general and not tied to a specific DL like ALC, though ALC is used for concrete examples.
  • Partially Covered: The ontology doesn't need to cover the entire vocabulary of the PS.

Key methodological contributions include:

  • Hybrid Rule Condition Checking: They propose a way to check rule conditions that uses the working memory directly for PS predicates (traditional approach) and DL entailment for DL predicates.
  • Traditional Rule Retraction: For rule effects, they opt for the traditional PS approach of simply removing atoms from the working memory. This simplifies things, avoids inconsistencies, and is always expressible, even if it doesn't always change the semantic truth value of a fact in the DL context.
  • Traditional Rule Execution: They stick to the traditional PS approach of firing rules if they syntactically change the working memory, rather than waiting for a semantic change. This is justified by the potential for semantically equivalent but syntactically different working memories when the ontology doesn't fully cover the PS vocabulary.
  • Fixed-Point Logic Embedding: The POS semantics is embedded into Fixed-Point Logic (FPL). This provides a formal, model-theoretic semantics and allows for reasoning about system properties like termination using logic entailment.

Datasets

Not discussed in the paper. The paper uses a running example of a research institute with staff and visitors to illustrate its concepts.

Key Insights and Ideas

  • Bridging the Semantic Divide: The core insight is that a hybrid approach is necessary. You can't simply force PS into a DL framework or vice-versa without losing crucial functionality or introducing complexity. POS offers a pragmatic way to combine them.
  • Pragmatic Retraction: The decision to use traditional PS retraction (syntactic removal) is a key pragmatic choice. While it might not always change the DL-implied truth of a fact, it keeps the system manageable and consistent, avoiding the complexities of DL ABox updates.
  • Independency of Concepts: The paper highlights the importance of treating certain PS concepts as "independent" from the ontology. This is crucial when the ontology doesn't fully specify relationships, allowing the PS to operate locally without being overly constrained by incomplete global knowledge.
  • Formalization for Analysis: Embedding the combined semantics into Fixed-Point Logic is a powerful move. It transforms the operational semantics of PS into a declarative one, enabling formal proofs and decidability results for properties like termination.

Key Results

  • POS Semantics: A novel semantics (POS) is defined for combining PS and DLs, addressing key challenges in rule condition checking, rule effects, and rule execution.
  • FPL Embedding: The POS semantics is embedded into Fixed-Point Logic, providing a model-theoretic foundation.
  • Decidability Results: The paper proves that checking termination for specific classes of combined systems – regular and light concrete PS – is decidable. This is a significant result, as termination is not generally decidable for production systems.

Limitations and Future Work

  • Specific DL: While the semantics is designed to be DL-independent, the paper uses ALC for concrete examples. Extending to more expressive DLs could be a future direction.
  • RIF-PRD Language: The authors plan to extend their work to cover the whole RIF-PRD (Rule Interchange Format - Production Rule Dialect) language, which includes more complex constructs than what was considered in this paper.
  • Inconsistencies: The paper acknowledges the challenge of handling inconsistencies that might be introduced by rule applications and plans to address this in future work.
  • Full RIF-PRD: The authors explicitly state their future plan to extend this work to cover the entire RIF-PRD language.

Summary updated

BibTeX
@inproceedings{rezk2011combining,
  author = {Martín Rezk and Werner Nutt},
  title = {{Combining Production Systems and Ontologies}},
  booktitle = {Rr},
  year = {2011},
  doi = {10.1007/978-3-642-23580-1\_26},
}
Figure