Introduction
In Knowledge Representation (KR), researchers often integrate Description Logics (DLs), which excel at defining hierarchies and ontologies, with Logic Programs (LPs), which are optimized for rules and exceptions. However, merging these into a "Hybrid Knowledge Base" frequently leads to contradictions, especially when aggregating data from heterogeneous sources.
Standard logic systems suffer from the "Principle of Explosion": if a single contradiction is detected (e.g., a sensor being both "on" and "off"), the system concludes that every possible statement is true, rendering it useless. This paper introduces two novel methods for paraconsistent reasoning that allow AI systems to draw sensible conclusions despite local inconsistencies, maintaining computational efficiency without requiring manual data correction.
Methodology
The authors' approach is built upon the MKNF (Minimal Knowledge and Negation as Failure) framework. They introduce two specific multi-valued semantics to handle contradictions locally.
5-Valued Semantics
This system utilizes five truth values: true (t), false (f), undefined (u), inconsistent (b), and undefined-false (uf). By isolating contradictions, the system allows the rest of the knowledge base to function normally.
6-Valued Semantics and Suspicious Reasoning
The 6-valued system introduces a sixth value, suspicious (s). This implements "Suspicious Reasoning," which tracks whether a conclusion was derived from a contradictory segment of the database. If a conclusion relies on "shaky" information, it is marked as suspicious rather than definitely true, providing a more nuanced layer of data integrity.
Efficiency via Fixpoint Algorithm
The core of the system's efficiency is a fixpoint algorithm. Rather than performing complex global consistency checks, the system iteratively updates the truth values of atoms until they stabilize. This ensures that if the underlying ontology is tractable, the entire hybrid system remains computationally efficient.
Key Insights and Results
- Maintained Tractability: The authors prove that adding paraconsistency does not increase the complexity class of reasoning. Computing the "well-founded" model remains in P (polynomial time) relative to the complexity of the underlying Description Logic.
- Faithfulness: The systems are "faithful," meaning that in the absence of contradictions, they produce the exact same results as standard, non-paraconsistent logic.
- Logical Soundness: The paper provides formal proofs that their fixpoint algorithms correctly represent the intended 5-valued and 6-valued models and generalize established standards like $WFSX_p$ and $ALC4$.
Limitations and Future Work
The current approach requires the hybrid knowledge base to be "DL-safe," a common restriction to ensure decidability. While the theoretical foundation is robust, future work involves high-performance software implementation, potentially via the NoHR (Nova Hybrid Reasoner) Protégé plug-in. Additionally, the authors suggest exploring "Suspicious Reasoning" within the ontology (DL) component itself, as the current 6-valued logic focuses primarily on the rules component.
Summary updated