Abstract:
"Traditional aspect-based sentiment analysis involves extracting themes from unstructured data and tagging each review with keywords linked to the themes. However, this approach assigns the overall polarity of a review to all associated themes, neglecting the nuanced sentiment for each theme. In Sri Lanka, where over 25,000 graduates enter the job market annually, there are no tools to help them assess company sentiment by specific aspects.
This study aims to develop a framework to analyze the sentiments of Sri Lankan companies across various aspects using deep learning models on online reviews. A total of 1,064 Glassdoor reviews were scraped for Sri Lanka’s Top 20 companies ranked by Brand Finance. Utilizing GPT-4o, a deep learning model, a novel method was implemented to annotate reviews at the aspect level. A hybrid scoring method was then introduced, balancing human and AI annotations.
The analysis resulted in 3,009 ratings for the 1,064 reviews, visualized through Power BI. Insights revealed Singer as the top company for career growth, Elephant House for work environment and wellbeing, Dilmah for corporate reputation and CSR, Distilleries Company of Sri Lanka for compensation, and Nestlé for leadership and management. The tool allows granular exploration of data by company, aspect, job title group, or industry sector.
This study highlights a critical limitation in traditional sentiment analysis, where reviews discussing both pros and cons are inaccurately summarized. With advancements in deep learning, it is now feasible to capture the nuanced sentiment of individual themes within a review."
