Abstract
This paper explores a bridge between the rigid, symbolic logic of ontologies and the fluid, uncertain world of statistical machine learning. The authors propose a new algorithm for semi-automatically learning ontologies using Probabilistic Description Logics ($cr\mathcal{ALC}$). By combining Inductive Logic Programming (ILP) with a Noisy-OR probabilistic classifier and the Expectation-Maximization (EM) algorithm, they demonstrate how to extract both logical structures and their associated probabilities from noisy, real-world datasets.
Introduction
Ontologies—formal representations of categories, properties, and relationships—are the backbone of the Semantic Web. However, building them manually is a burdensome and difficult task. While automated tools exist, most struggle with two primary challenges: expressivity and uncertainty. Many tools can only learn simple hierarchies, and standard Description Logics (DLs) are deterministic, failing to handle the "maybe" or "likely" scenarios common in real-world data. This research aims to fill this gap by creating a system that learns complex logical definitions while simultaneously calculating the probability that those definitions hold true.
Methodology
The proposed hybrid approach blends symbolic logic with statistical inference through several key components:
Logic Framework ($cr\mathcal{ALC}$)
The authors utilize $cr\mathcal{ALC}$, a specific probabilistic extension of Description Logic. This allows for "probabilistic inclusions," such as $P(\text{Student}|\text{Researcher}) = 0.4$, indicating a 40% probability that a researcher is also a student.
Structure Learning via ILP
To determine the logical "shape" of a concept (e.g., the definition of a "Professor"), the system uses Inductive Logic Programming (ILP) techniques inspired by the FOIL algorithm. It searches through candidate hypotheses using refinement operators to identify which logical constructs best fit the observed data.
Probabilistic Classification with Noisy-OR
To manage uncertainty during the search process, the authors employ a Noisy-OR classifier. This is a simplified Bayesian model that assumes different causes for an event act independently. It is computationally efficient and helps guide the search for the best logical structures by providing a probabilistic score for candidate definitions.
Parameter Estimation
Once a logical structure is identified, the Expectation-Maximization (EM) algorithm is used to fine-tune the probability values based on the observed data, ensuring the model accurately reflects the statistical properties of the dataset.
Case Study: The Lattes Platform
The algorithm was tested on the Lattes curriculum platform, a massive public repository maintained by the Brazilian government containing the CVs of thousands of researchers and students. Data was extracted from HTML pages to create a restricted relational database including concepts like "Researcher," "Publication," and "Organization," as well as roles like "advises" or "worksAt."
Key Insights and Results
- The Power of "OR": Traditional ILP often relies on strict conjunctions (AND). The authors found that the probabilistic Noisy-OR approach provides a more flexible "soft" logic that better handles the variations found in real-world CVs.
- Handling the Open World: Description Logics typically operate under the Open World Assumption (OWA). The authors adapted their learning routine to ensure that "unknown" instances do not unfairly penalize candidate hypotheses.
- Defining a Professor: The system successfully learned a definition for "Professor" involving being a person who has publications, advises students, or works at an organization. Given basic evidence, it assigned a probability of 0.68 to an individual being a professor.
- Detecting Duplicates: The algorithm learned that a duplicate publication is likely if it shares a similar title, year, and type. Title similarity alone yielded a probability of 0.77, while adding "Same Type" evidence pushed the probability to 0.99.
Conclusion
Preliminary results demonstrate that the algorithm can successfully learn complex definitions from noisy data. Future work will focus on improving scalability for larger domains, moving toward a fully automated pipeline from raw web data to probabilistic ontologies, and refining the operators used to navigate the logical search space.
Summary updated