Subject Area

Computer Science

Abstract

Commonsense reasoning has long presented a hurdle between conversational agents and their ability to naturally engage with humans in conversation, as the infinitely dimensional nature of a dialogue’s topics presents a significant reasoning challenge in the study of Natural Language Understanding (NLU). Such a system must conceivably be able to act as a generalizable system for evaluating and reasoning about commonsense statements, problems, and queries: in this way, the agent can attempt to quantify the reasonability of a given input. We attempt to address this through the integration of an explainable neuro-symbolic system that leverages Logical Tensor Networks (LTNs) and First-Order Logic (FOL), combined with sub-symbolic representations of topics and a dynamic baseline knowledge base (KB), to enable generalized reasoning over an unbounded domain of subjects presented in the form of natural sentences and achieve SOTA performance on Yale’s FOLIO v0.0 dataset.

Such a system is also capable of learning even during the inference of the model, and this being the case, we present a methodology through which the immense KBs associated with such a system might be efficiently evaluated to enable the timely training, inference, and growth of the KB. We also present a novel system for analyzing and controlling the quantifications and predicates learned by the agent, such that they are aligned with user intent, as well as preventing potentially unwanted changes of the KB and associated agent groundings and representations.

Degree Date

Spring 5-11-2024

Document Type

Thesis

Degree Name

M.S.

Department

Computer Science

Advisor

Dr. King-Ip Lin

Second Advisor

Dr. Eric Larson

Third Advisor

Dr. Ginger Alford

Number of Pages

69

Format

.pdf

Creative Commons License

Creative Commons Attribution-Noncommercial 4.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License

Available for download on Saturday, May 10, 2025

Share

COinS