PhD Thesis Proposal: Anthony Ragazzi

ZOOM LINK
Meeting ID: 918 8766 6056
Passcode: 458825

"A System for Automatic Ontology Generation Through Large Language Models and Inconsistency-Tolerant Reasoning"

Abstract

Ontologies, powerful tools employed to perform computer reasoning, have been successfully applied in a variety of applications.  Despite their utility, an important limitation hinders wider adoption. Since they are built upon logical formulations, they are also governed by logical rules.  As a result,  the possibility of logical contradictions presents a major hurdle that must be resolved. Designing ontologies that are consistent is costly and time consuming, and automatic methods lack the logical reasoning capabilities required to ensure coherence.  While this has been the subject of extensive research, current methods require information to be removed or revised to restore consistency.  As a result, users are forced to use the best available ontology for their application, which may not be an ideal representation of that domain.  There is a need for an ontology reasoning framework that is tolerant of conflicts.  Handling inconsistencies would enable the implementation of automatic ontology generation systems, making ontologies more accessible to a wider range of applications.  

The proposed work addresses these issues by presenting the following.  The first component is a theoretical framework for probabilistic ontologies.  This Bayesian framework allows for reasoning under uncertainty and incompleteness, as well as for the fusion of distinct and potentially conflicting ontologies.  Second, an ontology repair algorithm is presented which performs multiple repairs, aimed at including all relevant axioms without modification.  Lastly, we propose an automatic ontology generation procedure using large language models. Combined, the proposed work presents a system for automatic ontology generation that is designed to overcome challenges in consistency handling.  By proposing a reasoning and repair framework alongside a generation process we allow information to be extracted from the large language model without needing to navigate consistency constraints, which are handled by the other system components.

Thesis Committee

• Eugene Santos (Chair)
• George Cybenko
• Vikrant Vaze

Contact

For more information, contact Thayer Registrar at thayer.registrar@dartmouth.edu .