Cardiff University | Prifysgol Caerdydd ORCA
Online Research @ Cardiff 
WelshClear Cookie - decide language by browser settings

Deep learning approach to sentiment analysis in health and well-being

Zunic, Anastazia 2022. Deep learning approach to sentiment analysis in health and well-being. PhD Thesis, Cardiff University.
Item availability restricted.

[thumbnail of Anastazia Zunic PhD Thesis]
PDF (Anastazia Zunic PhD Thesis) - Accepted Post-Print Version
Available under License Creative Commons Attribution No Derivatives.

Download (2MB) | Preview
[thumbnail of Anastazia Zunic ORCA Form] PDF (Anastazia Zunic ORCA Form) - Supplemental Material
Restricted to Repository staff only

Download (117kB)


Sentiment analysis, also known as opinion mining, is an area of natural language processing which focuses on the classification of the sentiment that is expressed in a written document. Sentiment analysis has found applications in various domains including finance, politics, and health. This thesis is focused on sentiment analysis in the domain of health and well-being. An extensive systematic literature review was carried out to establish the state of the art in sentiment analysis in this domain. This systematic review provides evidence that the state-of-the-art results in sentiment analysis in the domain of health and well-being lags behind that in other domains. Additionally, it revealed that deep learning has not been used to classify the sentiment within the aforementioned domain. Furthermore, we performed a study and showed that the language that is used within the health and well-being domain is biased towards the negative sentiment. Aspect-based sentiment analysis refines the focus of sentiment analysis by classifying the sentiment associated with a specific aspect. Subsequently, we focus specifically on aspect-based sentiment analysis. To support it within the domain of health and well-being we created a dataset consisting of drug reviews, where the aspects were automatically annotated by matching concepts from the Unified Medical Language System. We have successfully shown that graph convolution can effectively utilise the context, represented with syntactic dependencies, to determine the intended sentiment of inherently negative aspects and consequently close the performance gap regardless of the domain. The advent of transformer-based architectures initiated a breakthrough in various tasks in natural language processing, including sentiment analysis. There-fore, we presented an approach to fine-tuning a transformer-based language model for the specific task of aspect-based sentiment analysis. The findings show the evidence that transformer-based models account for syntactic dependencies when classifying the sentiment of the given aspect.

Item Type: Thesis (PhD)
Date Type: Completion
Status: Unpublished
Schools: Computer Science & Informatics
Subjects: Q Science > QA Mathematics > QA76 Computer software
Date of First Compliant Deposit: 22 September 2022
Date of Acceptance: 15 September 2022
Last Modified: 23 Sep 2022 09:18

Actions (repository staff only)

Edit Item Edit Item


Downloads per month over past year

View more statistics