Personalized Reasoner Based on Belief Strengths of Information Sources

Shu-Bin Cai, Zhong Ming, Shixian LiAAAI 2008

Abstract

Traditional Description Logics (DLs) treat facts as universally true or false, failing to account for the subjective and often contradictory nature of information in open environments like the Semantic Web. This paper introduces Belief Lattice Description Logic (BLDL), a framework that personalizes reasoning by weighing information based on the "belief strength" of its source relative to the user. By utilizing a lattice-based truth system and dual negation, BLDL can resolve inconsistencies and provide context-aware answers without the logical collapse typical of classical systems.

Introduction

In the evolving landscape of the Semantic Web, information is rarely monolithic. Traditional AI reasoners operate under the assumption that facts are individual-independent—what is true for one agent must be true for all. However, real-world data is frequently subjective, context-dependent, and contradictory. This paper, presented at the 2008 AAAI Conference, proposes a "Personalized Reasoner" that acknowledges these nuances by calculating the "belief strength" of information sources based on their relationship to the questioner.

The Challenge of Inconsistency

Standard Description Logics (DLs) are the formal backbone of ontologies. While powerful, they are brittle in the face of contradiction. In classical logic, the "principle of explosion" dictates that from a single contradiction, anything can be proven, rendering the entire reasoning system useless. In an open internet environment where different ontologies (e.g., conflicting medical advice) may coexist, manually resolving every inconsistency is impossible. The authors aim to automate this resolution by prioritizing information according to the user's perspective.

Methodology: Belief Lattice Description Logic (BLDL)

The proposed framework, BLDL, extends standard Description Logic through a three-step personalized reasoning process:

  1. Identification: The system identifies both the information providers and the questioners using existing ontologies.
  2. Belief Strength Calculation: A "semantic distance" (ranging from 0 to 1) is calculated between the provider and the user. This distance is then mapped to a specific belief strength ($d_i$) within a Belief Lattice (BL).
  3. Reasoning with BLDL: The core engine uses multi-valued logic rather than binary true/false values.
Belief Lattice structures showing (a) a basic four-valued logic and (b) a multi-valued belief lattice with varying strengths $d_i$.

Dual Negation and Preferential Models

A key innovation in BLDL is the distinction between two types of negation:

  • Ontological Negation ($\neg$): Represents "it is not the case that P."
  • Epistemic Negation ($\sim$): Represents "it is known to be false that P."

This distinction allows the system to represent complex states like "unknown" or "contradictory" explicitly. The truth table for epistemic negation is defined within the lattice structure to maintain logical consistency across different belief strengths.

Truth table for epistemic negation ($\sim$) within the BLDL framework.

To handle contradictions, the reasoner seeks preferential models. Instead of failing when an inconsistency is found, the system identifies interpretations that minimize the set of epistemically inconsistent atoms. This "most consistent" possible world approach ensures the reasoner remains functional even in noisy environments.

Key Insights

  • Relativity of Truth: Truth is treated as a function of trust or proximity. A fact from a highly trusted source carries more weight ($d_i$) than one from a distant source.
  • Inconsistency Management: By focusing on minimization rather than elimination, BLDL provides a pragmatic path forward for reasoning with imperfect data.
  • Computational Feasibility: The authors adapt the standard "tableaux" algorithm for DL to work with multi-valued belief strengths. They argue that BLDL is as practical as standard DL; it only reaches worst-case complexity when no consistent model can be found.

Results and Complexity

The paper establishes that reasoning within BLDL is decidable, guaranteeing that the algorithm will eventually terminate with an answer. Furthermore, the authors provide a formal proof (Theorem 1) for the conditions of preferential unsatisfiability. Crucially, the time complexity of BLDL is shown to be on par with normal DLs (typically Exp-Time or N-Exp-Time), meaning the added personalization does not necessarily come at a prohibitive computational cost.

Limitations and Future Directions

Despite its theoretical strengths, several hurdles remain for practical implementation:

  • Distance Functions: The paper leaves the specific implementation of the "semantic distance" function as an open design choice.
  • Logic Expressivity: The current framework focuses on simpler logics (SI and SHI). Extending these principles to more expressive logics like SHIQ or SHOIQ, which are common in modern web standards, remains a subject for future research.
  • Empirical Validation: As a theoretical work, the paper lacks real-world benchmarks or user studies to confirm that the personalized outputs align with human expectations.

Summary updated

BibTeX
@inproceedings{cai2008personalized,
  author = {Shu-Bin Cai and Zhong Ming and Shixian Li},
  title = {{Personalized Reasoner Based on Belief Strengths of Information Sources}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2008},
}
Figure