Natural Language Understanding and Commonsense Reasoning Using Answer Set Programming and Its Applications
Abstract
Abstract
The goal of natural language understanding (NLU) research is to build systems that can exhibit human-like intelligent behavior. Intelligent behavior in humans includes both learning
and reasoning. With the dramatic success of Machine Learning (ML) algorithms, researchers
have created several state-of-the-art NLU applications that are purely based on learning with
no, or shallow, reasoning capabilities. Also, machine learning-based NLU applications behave
like a black-box, which means they are unable to provide justification for responses they generate, even if these responses have high accuracy. That is, these applications are not explainable.
A crucial component of the human thought process—commonsense reasoning—is missing in
almost all of these machine learning-based solutions. This dissertation shows how default logic
can be employed for commonsense reasoning and building efficient logic-based NLU applications. Our research leverages the declarative logic-based paradigm of Answer Set Programming
(ASP) to efficiently represent knowledge—essential for any explainable NLU application—as
a default logic theory. With the help of goal-directed ASP solvers such as the s(ASP) and
s(CASP) systems, our applications are capable of providing (natural language) justification
for any computed action. We show that our work is not just limited to purely NLU applications
such as automated natural language question-answering (QA) systems or automated chatbots,
but also to the field of visual question answering (VQA) and text-based games.