Uncovering Hidden Issues in Audit Findings Through LDA-Based Topic Modeling

Abstract

Academic audit reports play an important role in assessing and monitoring the quality of higher education. However, most of these reports are arranged in an unstructured narrative descriptive form, making it difficult to analyze systematically and consistently, especially if done manually. This poses a challenge for auditors and decision makers in identifying patterns of findings and quality issues efficiently. This study aims to apply and evaluate the Latent Dirichlet Allocation (LDA) method in extracting keywords and abstracting main topics from academic audit report texts. The dataset was obtained from the Quality Management System (SIMUTU) of Surabaya State University, which includes hundreds of audit finding descriptions from various faculties over the past three years. The methodology used includes text preprocessing stages using tokenization, stopword removal, and stemming techniques, followed by topic modeling using LDA. Evaluation was carried out quantitatively using a coherence score to assess topic quality, and qualitatively through visualization of results in the form of word clouds and pyLDAvis. The results showed that the LDA model was able to produce meaningful, representative, and relevant topics in the context of academic quality, such as document management, lecturer involvement, and implementation of learning evaluations. Manual validation by internal quality experts showed that the generated topics can help in understanding audit findings trends more quickly and objectively. Thus, LDA has proven to be effective as an approach to extracting important information from unstructured audit reports and has great potential to be integrated into data-driven quality dashboard systems to support more informed and evidence-based decision making.