Semantic Query Extension through Probabilistic Description Logics

José Eduardo Ochoa Luna, Kate Revoredo, Fábio Gagliardi CozmanURSW 2010 · OpenReview

Introduction

Standard keyword-based information retrieval systems are often limited by their literal interpretation of queries. A search for "Publications about probabilistic fish" might overlook a highly relevant paper titled "A new kind of aquatic vertebrate with probabilistic processing," even if the content is exactly what the user needs. While integrating ontologies—structured knowledge graphs—can bridge this semantic gap, it introduces the challenge of uncertainty. Not every linked concept in an ontology is equally relevant to a specific query. This research proposes a framework to mathematically model this uncertainty using probabilistic description logics, specifically Credal $\mathcal{ALC}$ (CR$\mathcal{ALC}$), to calculate the probability of relevance and rank search results accordingly.

Methodology

The core of the proposed framework is the use of CR$\mathcal{ALC}$, which extends traditional description logics by allowing for probabilistic inclusions. While standard logics define strict rules (e.g., "Every Professor is a Researcher"), CR$\mathcal{ALC}$ allows for nuanced statements such as "The probability that a Researcher is a Professor is 0.4."

The search framework operates through a three-step pipeline:

  1. Search Procedure: An initial keyword search is performed on indexed documents, which are then mapped to concepts and roles within a pre-learned probabilistic ontology.
  2. Query Extension: These initial hits serve as hard evidence to instantiate a Relational Bayesian Network (RBN). Through probabilistic inference, the system calculates the likelihood of related, non-evidence concepts. Those exceeding a defined probability threshold are selected as query extensions.
  3. Ranking Procedure: Documents associated with these inferred concepts are retrieved, merged with the original results, and ranked based on their computed probabilities.
Probabilistic ontology definitions and inclusions used to model the research domain.

The authors demonstrated this approach using the Lattes Curriculum Platform, a structured repository of data on Brazilian researchers. A subset of 1,964 documents was used to automatically learn the probabilistic terminology and evaluate the system's performance.

Results and Implementation

A key feature of this framework is the decoupling of raw text documents from the ontology, linked only via text indexes. This allows the probabilistic ontology to serve as a flexible template. Keyword hits act as triggers; for example, a keyword appearing in a publication title provides evidence for a specific node in the Bayesian network, which then propagates to elevate the probability of related nodes, such as co-authors or advisors.

The Probabilistic Semantic Search Engine interface showing results for a query on "Bayesian networks."

Qualitative evaluation across 20 random research topics showed that the system consistently identified highly relevant researchers who did not explicitly match the initial keywords but were strongly connected through the relational network. The probabilistic ranking mechanism achieved a 99% positive analysis in testing, successfully prioritizing results based on the cumulative strength of their semantic relationships.

Conclusion

While the framework provides a robust mathematical foundation for semantic search, computational scalability remains a challenge. Exact probabilistic inference becomes intractable as ontologies grow and the number of instances increases. Future work will explore approximate inference techniques, such as variational methods or first-order loopy belief propagation, to enable web-scale application. Additionally, the field requires more standardized benchmark datasets to objectively evaluate relational semantic search systems.

Summary updated

BibTeX
@inproceedings{luna2010semantic,
  author = {José Eduardo Ochoa Luna and Kate Revoredo and Fábio Gagliardi Cozman},
  title = {{Semantic Query Extension through Probabilistic Description Logics}},
  booktitle = {URSW},
  year = {2010},
  url = {https://openreview.net/forum?id=SXiEtWk13B4},
}
Figure