BizSA: webapp to visualise ABSA of customer reviews

• Main Author: Liew, Shaw Kee
• Other Authors: Erik Cambria
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/166649
• Institution: Nanyang Technological University

ABSA is a popular field of study in NLP. It is applicable and useful in many contexts. One such context is the usage of it for businesses to identify the sentiment of the customers with respect to the different aspects of their products. Although there are many models proposed to tackle the task of ABSA, most require labeled training data to be inputted on the part of the user for high accuracy. Considering that some companies might not have the time to label the data, or they simply do not have access to large amount of data, this paper aims to provide a toolkit called BizSA that helps users to perform ABSA without the need for training data using GPT-3. With prompt design of GPT-3, BizSA allows users to easily customise some examples for the prompt input. This enables users to specify some of the aspect terms and aspect categories specific to the domain of their datasets. Although BizSA provides users with a simple and efficient way to analyse and visualise data, it was observed that GPT-3 had trouble grouping similar terms together through experiments. Hence, it tends to generate more terms than it should, which could cluster graphs and make visualisation hard for users.