A Declarative Approach to Data-Driven Fact Checking

Julien LeblayAAAI 2017 · DOI

Introduction

Automated fact-checking often fails by reducing complex, nuanced claims to binary "true" or "false" labels. In reality, the truth of a statement is frequently contingent on its context—specifically when it was said, who said it, and what underlying belief system is being applied. This paper introduces a formal, declarative framework that treats context as a first-class citizen, enabling sophisticated queries over incomplete and uncertain data.

Methodology: The Contextual Chase

The proposed framework integrates three distinct logical and probabilistic components to handle the nuances of real-world data:

  1. Datalog±: A logic language used for ontological reasoning, allowing the system to infer missing information using predefined rules.
  2. Lattice-based Annotations: A mathematical structure that organizes contexts (such as time intervals or source sets) hierarchically, allowing facts to be annotated with their specific validities.
  3. Markov Logic Networks (MLNs): These handle "soft" rules by assigning weights to formulas, allowing the system to calculate the probability of a claim within a specific context.

The core of the system is the Contextual Chase, an algorithm that propagates these annotations through a knowledge base. To ensure efficiency, the paper introduces a semi-naïve evaluation strategy for "Scope Queries," which determine if a claim holds in a specific context.

Algorithm 1: Semi-Naïve evaluation for scope queries.

Key Results and Expressivity

The framework's effectiveness is demonstrated through its ability to handle multi-dimensional contexts. For instance, in a scenario involving political stances, the system can track how confidence in a claim changes as different sources are combined or as time progresses. This allows for "Support Queries," which identify all contexts where a claim's confidence exceeds a certain threshold.

Visualization of claim scores across source combinations (i) and temporal intervals (ii).

Crucially, the author proves that query answering within this framework is PTIME-complete in data complexity under guarded TGDs. This theoretical result ensures that the approach is computationally feasible for practical applications on reasonably sized databases. Unlike standard logic engines that might fail upon encountering a contradiction, this system tracks conflicts as specific annotations, providing transparency into why and where information diverges.

Conclusion and Future Directions

By formalizing context through lattices and MLNs, this research provides a robust foundation for next-generation fact-checking tools that move beyond binary truth. While currently limited by the scalability of large context spaces and the need for manual weight assignment, future work aims to incorporate automated weight learning and approximate methods for web-scale deployment.

Summary updated

BibTeX
@inproceedings{leblay2017declarative,
  author = {Julien Leblay},
  title = {{A Declarative Approach to Data-Driven Fact Checking}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2017},
  doi = {10.1609/aaai.v31i1.10492},
}

Loading related content...

Figure