Inconsistency Tolerant Reasoning in Knowledge Bases

32 concepts30 papers3 proposals

In classical logic, a single contradiction destroys the entire system. This is known as the principle of explosion (ex falso quodlibet): from a contradiction, anything follows. If a knowledge base contains both $A$ and $\neg A$, standard refutation-based theorem provers suffer a "logical meltdown." They become incapable of producing meaningful results because every possible statement suddenly becomes provably true, a foundational hurdle highlighted by Didier Dubois et al. in Theorem Proving Under Uncertainty - A Possibility Theory-based Approach.

Yet, real-world data is inherently messy. Whether we are integrating distributed sensors, merging heterogeneous databases (Laurence Cholvy in Proving Theorems in a Multi-Source Environment), extracting web-scale relations (Alfio Gliozzo et al. in Populating Web-Scale Knowledge Graphs Using Distantly Supervised Relation Extraction and Validation), or parsing cyber threat intelligence from multiple vendors (Yuqiao Meng et al. in Uncovering Vulnerabilities of LLM-Assisted Cyber Threat Intelligence), contradictions are unavoidable.

To build intelligent systems that do not collapse under real-world noise, researchers developed frameworks for inconsistency-tolerant reasoning. Instead of forcing a knowledge base to be perfectly consistent, these methods allow systems to reason through and around contradictions.

Coherence via Maximal Consistent Subsets

One intuitive way to handle an inconsistent knowledge base is to isolate the contradictions by breaking the database down into its largest possible uncontradicted chunks. These are called Maximal Consistent Subsets (MCS).

Imagine a knowledge base $\mathcal{K} = \{A, \neg A, B\}$. The entire set is inconsistent. However, we can extract two maximal consistent subsets: $\{A, B\}$ and $\{\neg A, B\}$. Notice that $B$ remains present in both. If we only infer conclusions that are true across these subsets, we can safely conclude $B$ without letting the contradiction between $A$ and $\neg A$ poison our reasoning.

Sébastien Konieczny et al. in Rational Inference Relations from Maximal Consistent Subsets Selection formalize this by selecting specific subsets of MCSs using monotonic criteria, ensuring that the resulting inferences remain logically rational. Similarly, Salem Benferhat et al. in How to Infer from Inconsisent Beliefs without Revising? demonstrate how systems can maintain logical sanity without undergoing the computationally expensive process of constantly revising or deleting beliefs.

This subset-selection paradigm is highly flexible. For instance, a unified framework detailed in Reasoning under Inconsistency Based on Implicitly-Specified Partial Qualitative Probability Relations: A Unified Framework allows users to implicitly specify preferences via constraints to isolate preferred consistent subsets efficiently. By focusing algorithms on identifying and correcting the underlying sources of these conflicts, as discussed by the authors of Reasoning About Inconsistent Formulas, systems can pinpoint exactly where a premise fails.

Graded Certainty and Possibilistic Logic

Not all pieces of information are created equal. If a highly reliable sensor contradicts a noisy one, we should not treat their claims with equal weight. Possibilistic logic addresses this by attaching certainty levels or priority degrees to classical logic formulas.

When merging prioritized knowledge bases, a naive approach might discard all lower-priority information the moment a high-priority contradiction occurs. This is known as the drowning problem. To circumvent this, Guilin Qi in A Model-based Approach for Merging Prioritized Knowledge Bases in Possibilistic Logic proposed an iterative, layer-by-layer merging operator that resolves conflicts at the highest priority levels while preserving safe, lower-priority data.

As surveyed by Didier Dubois et al. in 40 Years of Research in Possibilistic Logic - a Survey, this mathematical framework provides a robust foundation for quantifying inconsistency rather than merely treating it as a binary failure state.

Argumentation and Strategic Dialogue

Another powerful perspective treats inconsistency as a live debate. Instead of viewing a knowledge base as a static set of truths, we can view it as a pool of competing arguments. In this view, an inference is valid if the arguments supporting it can withstand the counterarguments.

The authors of NONMONOTONIC REASONING and Osamu Imatcht and Yuji established a formal bridge between these argumentation-based methods and coherence-based subset methods, proving that they are often mathematically equivalent ways to handle uncertainty and contradiction.

To make these systems practical, Leila Amgoud et al. in Compilation of Logical Arguments introduced techniques to "compile" arguments by stripping away irrelevant information, preventing the system from missing intuitive conclusions when an unrelated contradiction exists elsewhere in the knowledge base. Furthermore, argumentation can be framed dynamically: Yakoub Salhi in A Game-Theoretic Perspective on Inconsistency Handling modeled consistency restoration as a strategic game between a Proponent and an Opponent, using information entropy to measure the exact "effort" required to resolve a conflict.

This dialectical approach naturally extends to decision-making under uncertainty, as shown by Leila Amgoud et al. in Using Arguments for Making Decisions: A Possibilistic Logic Approach, where pros and cons are weighed to navigate conflicting goals and knowledge bases simultaneously.

Modern Applications: From Description Logics to LLM Agents

While early research focused primarily on propositional logic, modern applications demand inconsistency tolerance in highly complex, large-scale environments.

Semantic Web and Ontologies

In semantic web applications, large assertional datasets often clash with fixed ontological rules. The authors of Explaining Inconsistency-Tolerant Query Answering over Description Logic Knowledge Bases designed SAT-solver-backed frameworks to not only answer queries over inconsistent Description Logic bases but also provide users with clear, localized justifications pinpointing exactly which conflicting assertions caused the issue. Similarly, managing noisy extracted information remains a core challenge in automatic knowledge base construction for complex domains, as highlighted by Niranjan Balasubramanian et al. in A study of the knowledge base requirements for passing an elementary science test.

Interactive Robotics

In robotics, managing noisy inputs is critical for real-time operation. Marcelo Quinderé et al. in An information state based dialogue manager for a mobile robot implemented a Knowledge Acquisition and Management module that allows mobile robots to parse and reason through contradictory user statements during live dialogue without crashing the robot's internal world model.

Large Language Models and Multi-Agent Systems

The rise of Large Language Models (LLMs) has brought these classical challenges back to the forefront. LLMs frequently generate or encounter contradictory data when processing multi-source information.

Key Foundational Work

Theorem Proving Under Uncertainty - A Possibility Theory-based Approach (1987) — contributed 1 concept elements to this direction.
A Process for Evaluating Tree-Consistency (1980) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_inconsistent_knowledge_bases_f297a0a3,
  title = {inconsistent knowledge bases},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/inconsistent-knowledge-bases/rdr_f297a0a369a1410fb53ccdaed1cc6bd1},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_inconsistent_knowledge_bases_f297a0a3_v1,
  title = {inconsistent knowledge bases},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/inconsistent-knowledge-bases/rdr_f297a0a369a1410fb53ccdaed1cc6bd1/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake