Inconsistency Tolerant Reasoning for Knowledge Bases

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2133 papers152 conceptsfrozen 2026-10-04

In classical logic, a single contradiction is a death sentence for a reasoning system. This is due to the principle of explosion (ex falso quodlibet), a formal rule stating that from a contradiction, any statement can be proven true. If a database contains both "The sun is a star" and "The sun is not a star," a classical reasoner can logically conclude that "The moon is made of green cheese."

As Frank van Harmelen argues in Where Does It Break? or: Why the Semantic Web Is Not Just "Research as Usual", this brittleness is incompatible with the real world. Modern AI must ingest data from diverse, decentralized, and often conflicting sources—ranging from legal norms that clash (Yiwei Lu) to noisy sensor data in image labeling (Stefan Scheglmann). To build robust systems, researchers have developed frameworks that can "tolerate" inconsistency without collapsing into triviality.

Isolating the Explosion

The most direct way to handle contradictions is to change the underlying logic so that the principle of explosion no longer applies. These are known as paraconsistent logics.

Early work by João P. Martins and Stuart C. Shapiro on the Multiple Belief Reasoner (MBR) introduced the idea of tracking the "pedigree" of every belief. By tagging each fact with its source or derivation path, the system can reason within "belief spaces." If a contradiction is found, it is localized to a specific context rather than infecting the entire knowledge base. Fangzhen Lin furthered this in Reasoning in the Presence of Inconsistency by proposing a resolution principle that stays "nontrivial" even when premises conflict, allowing the system to provide useful answers for unrelated queries.

Repairing the Knowledge Base

Instead of changing the logic, another school of thought focuses on "repairing" the data. When a knowledge base contradicts its own rules (ontologies), we can look for maximal consistent subsets—the largest possible versions of the data that don't contain any contradictions.

Gerhard Brewka pioneered this with Preferred Subtheories, where premises are ranked by reliability. If two facts conflict, the system simply ignores the less reliable one. This "repair" approach has been scaled to massive datasets. For instance, Meghyn Bienvenu and colleagues developed CQAPri, a tool that uses SAT solvers to answer queries over inconsistent Description Logic bases by prioritizing more reliable data sources.

A more cautious version of this is Intersection ABox Repair (IAR) semantics. As explored by Thomas Lukasiewicz and Maria Vanina Martinez, IAR semantics only considers a statement true if it holds across all possible repairs. This "skeptical" approach ensures that the AI never provides an answer that depends on a disputed fact.

Reasoning as Argumentation

A third perspective treats logical reasoning as a form of debate. In Formal Argumentation in Symbolic AI, Markus Ulbricht 0001 describes how contradictions can be modeled as "attacks" between arguments.

In this framework, a conclusion is accepted not because it is "true" in an absolute sense, but because it belongs to a set of arguments that can defend themselves against all counter-arguments. This is particularly useful in domains like law, where Ofer Arieli and others use "status annotations" (Accepted, Rejected, Introduced) to show exactly why a specific rule was followed or ignored in the face of a conflict (Defeasible Normative Reasoning).

Quantifying the Messiness

Finally, some researchers argue that the binary "consistent vs. inconsistent" view is too narrow. Neutrosophic logic, introduced by Florentin Smarandache, models truth, falsity, and indeterminacy as three independent variables (Interval Neutrosophic Sets and Logic). This allows a system to say, "This statement is 70% true, 40% false, and 20% indeterminate," providing a mathematical way to handle over-defined or contradictory states.

Similarly, Alexander G. Gray and the team behind Logical Neural Networks use truth-value bounds. Instead of a single number, they track a range (e.g., [0.2, 0.8]). If the lower bound exceeds the upper bound, the system detects a contradiction but uses the "gap" to guide the learning process, effectively turning logical failures into signals for model improvement.

By moving away from the rigid requirements of classical logic, these diverse approaches allow AI to function in the "messy" reality of human knowledge, where contradictions are not errors to be deleted, but complexities to be managed.

Papers in this version

Showing 60 of 133 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 150 papers
  2. Version 1 · 2026-10-04 · 133 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_principle_of_explosion_badd6c17,
  title = {principle of explosion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/principle-of-explosion/rdr_badd6c177c764964932faa3282729e0d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_inconsistency_tolerant_reasoning_for_kno_badd6c17_v1,
  title = {Inconsistency Tolerant Reasoning for Knowledge Bases},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/inconsistency-tolerant-reasoning-for-knowledge-bases/rdr_badd6c177c764964932faa3282729e0d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}