SAMS Subject Classification: 17, 26

With the increasing volume and complexity of user-generated content shared via the Internet, the need has arisen for automated methods capable of extracting meaningful insights from unstructured text data. Sentiment analysis is a form of text analysis involving the process of computationally identifying the polarity of an opinion expressed by an author of a given piece of text. While much of the existing work in this field focusses on document-level or sentence-level analysis, in which an entire document or sentence, respectively, is viewed as a single information unit and is assumed to contain a maximum of one expression of sentiment, aspect-based sentiment analysis involves a more fine-grained approach, facilitating the discovery of any number of topics, and the sentiment polarities towards these topics, present in a document of text data [1]. The most promising approaches to aspect-based sentiment analysis to date are those based on supervised machine learning. Many of the methodologies in the literature are, however, focussed on the application of specific machine learning models, on only specific sub-tasks of the problem, or on a specific domain of application.

In this presentation, a generic framework for aspect-based sentiment analysis is proposed, the aim of which is to guide a user through the process of gaining insights from an unstructured text data set from any domain, and by utilising any appropriate machine learning models. The generality of the framework is demonstrated through its application to benchmark data sets from the literature, as well as on a case study involving a South African retail bank.

[1] B. Liu, Sentiment analysis: Mining opinions, sentiments, and emotions (\(2^{nd}\) Edition), Cambridge University Press, 2020.