Year: 2025
Type: dissertation
Abstract: The aim of this thesis is to improve upon the current state of transfer learning technology in text based problems, with a specific focus on sentiment analysis. This improvement is in the form of increased distance between training and target domains. It proposes a method for defining inter-domain distance. Then examines the shortcomings of classical transfer learning methods, and proposes two novel approaches. The first `TransferGAN' extends on the use of Adversarial Learning Techniques for transfer learning, and the second demonstrates how Grammatical Evolution can be used to optimize existing sentiment analysis techniques to better suit transfer learning between dissimilar domains. Both of these methods demonstrate improvement over the comparison systems for transfer learning sentiment analysis with a high inter-domain distance. Finally these systems are demonstrated to be effective against the Motivational Interview dataset (A dataset of clinical conversations between diabetes patients and clinicians), which was the primary motivator for this work. This work could be further extended by field-testing of the systems against real world usage of the Motivational Interviews, and by working with clinicians to refine the system for automatic assessment of the Interviews.
Author: Coop, Guy
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7135160232