Bo Huang MSc Defence
Date and Time
Location
Details
MSc Thesis Title: Comparative Study of Deep Learning Models for Sentiment Analysis
Abstract: The rapid expansion of online applications like social media apps and e-commerce
websites has led to a large volume of reviews about different subjects, products, and
services. Sentiment analysis, which is a crucial area of study in natural language
processing, aims to classify the sentiments of these reviews so that the feedback can
be valuable to companies, governments, and individuals in making informed decisions
based on the information gathered about the public opinions. This thesis presents a
comprehensive empirical study that investigates and compares multiple deep learning
approaches for sentiment analysis, including Convolutional Neural Networks (CNNs),
and variants of BERT-based models, which include pre-trained “bert-base-uncased”
and “roberta-case” models, hand-craft prompts, adaptive prompts, and hybrids of
these prompts. Results show that pre-trained “bert-base-uncased” and “roberta-case”
models outperform traditional CNNs, and prompt-based methods offer promising
results with reduced computational costs. Moreover, dataset characteristics, such
as input length, class distribution, and dataset structure, significantly impact the
performance of BERT-based models. Lastly, practical guidelines are proposed for
selecting appropriate models based on dataset characteristics.
websites has led to a large volume of reviews about different subjects, products, and
services. Sentiment analysis, which is a crucial area of study in natural language
processing, aims to classify the sentiments of these reviews so that the feedback can
be valuable to companies, governments, and individuals in making informed decisions
based on the information gathered about the public opinions. This thesis presents a
comprehensive empirical study that investigates and compares multiple deep learning
approaches for sentiment analysis, including Convolutional Neural Networks (CNNs),
and variants of BERT-based models, which include pre-trained “bert-base-uncased”
and “roberta-case” models, hand-craft prompts, adaptive prompts, and hybrids of
these prompts. Results show that pre-trained “bert-base-uncased” and “roberta-case”
models outperform traditional CNNs, and prompt-based methods offer promising
results with reduced computational costs. Moreover, dataset characteristics, such
as input length, class distribution, and dataset structure, significantly impact the
performance of BERT-based models. Lastly, practical guidelines are proposed for
selecting appropriate models based on dataset characteristics.