Towards a Policy Language for Managing Inconsistency in Multi-Context Systems

Thomas Eiter, Michael Fink, Giovambattista Ianni, Peter SchüllerLog-IC@LPNMR 2011 · OpenReview

Multi-Context Systems (MCSs) provide a robust framework for integrating heterogeneous knowledge bases through "bridge rules." However, as these systems scale and link disparate data sources—such as hospital databases, lab results, and medical ontologies—maintaining consistency becomes a significant challenge. This paper introduces the Inconsistency Management Policy Language (IMPL), a declarative framework designed to formalize and automate the resolution of logical conflicts within MCSs.

The Challenge of Inconsistency in MCSs

In modern AI applications, data is rarely centralized. A hospital information system, for instance, may link a patient database, laboratory results, and a medical ontology. These independent contexts are connected via bridge rules—logical links that allow information to flow between systems. Inconsistency arises when these rules produce conflicting conclusions, such as a patient being prescribed a medication that conflicts with a newly recorded allergy in a separate department.

While existing research has focused on diagnosing these inconsistencies, there has been a lack of a general-purpose, programmable strategy for managing them. Some conflicts are minor and can be resolved automatically, while others—particularly in high-stakes domains like medicine—require expert human intervention. IMPL addresses this gap by providing a formal, programmable strategy for handling diverse types of inconsistencies.

The IMPL Framework

IMPL is a logic-based language for specifying management policies. Inspired by Answer Set Programming (ASP), it allows developers to write high-level "if-then" rules to dictate how the system should respond to specific states of inconsistency. The language utilizes three primary components:

  • Reserved Predicates: These function as sensors, providing the policy with metadata about the system's state. For example, explNeed identifies bridge rules involved in a conflict, while diag identifies minimal sets of changes (diagnoses) that could restore consistency.
  • Action Atoms: These are the actuators of the system. They trigger specific operations, such as @delRule (deleting a bridge rule), @addRule (adding a rule), or @guiSelectMod (prompting a user to select a resolution via a graphical interface).
  • Three-Step Semantics: The execution of an IMPL policy follows a rigorous logical process:
  • Action Determination: The system calculates the "answer sets" of the policy to determine which actions are triggered.
  • Effect Determination: The actions are executed, which may include non-deterministic steps like waiting for user input.
  • Effect Materialization: The resulting changes are applied to the MCS to restore a consistent state.

Key Insights: Human-in-the-Loop and Separation of Concerns

A central contribution of IMPL is its treatment of human interaction as a first-class citizen. By incorporating user-selection actions into the logic, the system can defer complex decisions to experts. This "human-in-the-loop" approach ensures that automated repairs do not lead to unsafe or incorrect system states in sensitive environments.

Furthermore, IMPL promotes a "separation of concerns" by decoupling the analysis of a conflict (finding the diagnosis) from the management of that conflict (deciding the repair strategy). This modularity allows for more flexible and maintainable system architectures. The authors also propose an iterative methodology where the system fixes what it can automatically, re-evaluates the state, and then requests help if conflicts persist.

Feasibility and Future Work

The authors demonstrate the expressiveness of IMPL through a detailed hospital information system example, showing that the language can replicate complex repair strategies that previously required bespoke, hard-coded implementations. The language is practically feasible through an integration path with the acthex formalism, an extension of ASP that supports external actions.

Despite its strengths, the current iteration of IMPL focuses on modifying bridge rules rather than the internal data of the underlying knowledge bases. Future work aims to expand the language's reach to internal data and to implement formal termination guarantees for iterative repair processes to prevent potential infinite loops during automated resolution.

Summary updated

BibTeX
@inproceedings{eiter2011towards,
  author = {Thomas Eiter and Michael Fink and Giovambattista Ianni and Peter Schüller},
  title = {{Towards a Policy Language for Managing Inconsistency in Multi-Context Systems}},
  booktitle = {Log-IC@LPNMR},
  year = {2011},
  url = {https://openreview.net/forum?id=ckOIuLjiBV},
}
Figure