Temporal ABox Cleaning in TDL-Lite

Sabiha Tahrat, Salima Benbernou, Mourad OuziriDescription Logics 2021 · OpenReview

TL;DR

This paper tackles the challenge of inconsistent data in temporal knowledge bases (KBs) by proposing a novel approach to detect and repair these inconsistencies. They introduce a method to identify conflicting temporal assertions and then repair them by removing the "least valuable" assertions, prioritizing temporal relevance.

What Problem Does This Paper Address?

In many real-world applications, data is not static but evolves over time. Knowledge bases (KBs) are used to represent this information, but temporal data can become inconsistent due to updates, errors, or the inherent nature of time. For instance, a person might be recorded as "a child" at one time and "an adult" at a later time, which is perfectly fine. However, if the KB contains conflicting assertions like "John is a child at time 5" and "John is an adult at time 5," this represents an inconsistency. Traditional DL-Lite, a popular knowledge representation language, struggles to handle these temporal inconsistencies effectively. This paper focuses on a temporal extension of DL-Lite (TDL-Lite) and aims to provide a robust mechanism for detecting and repairing such temporal data inconsistencies.

Methods

The authors propose a two-pronged approach:

  1. Inconsistency Detection: They leverage an equi-satisfiable reduction from TDL-Lite to standard DL-Lite. This clever trick allows them to use existing, powerful DL-Lite reasoners (like Pellet) to pinpoint the exact set of inconsistent assertions within the temporal KB. They visualize these conflicts using an "inconsistency graph," which helps understand the relationships between conflicting assertions.
  2. Temporal Repair: Once inconsistencies are identified, they propose a "best temporal repair" strategy. This goes beyond simply removing conflicting assertions. They introduce a "temporal weight" for each assertion, considering its timestamp and whether it's associated with a "rigid predicate" (assertions that should ideally remain true over time). The repair process prioritizes removing assertions with lower temporal weights, aiming to preserve more recent or temporally significant information. They also define an "inconsistency degree" for each assertion, indicating how many conflicts it's involved in, and use this to guide the removal process when temporal weights are equal.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Bridging Temporal and Non-Temporal Reasoning: The reduction to DL-Lite is a key insight, enabling the reuse of mature reasoning technologies for temporal data.
  • Beyond Simple Deletion: The concept of "best temporal repair" is crucial. It acknowledges that not all inconsistencies are equal and that temporal context matters when deciding what to remove.
  • Temporal Weight as a Heuristic: Assigning a temporal weight based on timestamps and rigid predicates provides a principled way to prioritize assertions during repair.
  • Inconsistency Graph for Visualization: Visualizing conflicts helps in understanding the nature and scope of inconsistencies.

Key Results

The paper presents a formal framework for detecting and repairing temporal ABox inconsistencies in TDL-Lite. While specific quantitative results are not detailed in this excerpt, the methodology provides a clear path for implementing and evaluating such a system. The authors demonstrate how their approach can distinguish between different repair scenarios, as illustrated in Figure 1, where removing different assertions leads to temporally more sensible outcomes.

Limitations and Future Work

The authors acknowledge that this is an "exploration" and a "first step." Future work includes:

  • Investigating other temporal DL-Lite logics: Exploring logics that are rewritable to First-Order Logic.
  • Practical Implementation and Scalability: Building a repair framework and evaluating its performance on real-world data.
  • Metric Temporal Operators: Incorporating metric temporal operators (which allow specifying time intervals) into the TDL-Lite language for more expressive repair capabilities.
  • Data-driven or User-guided σ: Exploring how to set the "predicate lifespan" (σ) parameter, which influences the temporal weight of rigid predicates, either automatically from data or through user guidance.

Summary updated

BibTeX
@inproceedings{tahrat2021temporal,
  author = {Sabiha Tahrat and Salima Benbernou and Mourad Ouziri},
  title = {{Temporal ABox Cleaning in TDL-Lite}},
  booktitle = {Description Logics},
  year = {2021},
  url = {https://openreview.net/forum?id=nAKbTDJQGz},
}
Figure