Abstract
Topic modeling, particularly latent Dirichlet allocation (LDA), is widely recognized as a valuable technique for identifying key topics and trends across dynamic content in various fields. LDA’s strength lies in its ability to efficiently capture emerging themes from large text corpora, making it a popular choice for categorization. It facilitates the automation of report reviews, assisting in corporate evaluations and management assessments by uncovering key trends and topics with minimal manual intervention. However, our analysis of sustainability within the corpora of SGX-listed companies reveals limitations when using LDA on sparse data. Specifically, the dynamic LDA approach (dynamic topic modeling, or DTM), applied to an 11-year dataset of annual reports, struggles to detect the rise of sustainability as a significant corporate focus following policy changes. Despite the mandate for sustainability reporting, actual engagement with sustainability issues within these reports remains limited, i.e., highlighting the need for substantial improvements in how companies address sustainability topics.