<![CDATA[Digital transformation in higher education has increased the need for intelligent systems capable of organizing and classifying institutional correspondence efficiently, particularly when labeled training data are limited. This study develops and evaluates an Indonesian academic-correspondence classification system using SetFit within an efficient few-shot learning framework. A quantitative experimental design was applied to 145 official PDF documents collected from five units at Universitas Islam Sultan Agung. Native digital documents were extracted using PyMuPDF, whereas scanned documents were processed using Doctr OCR, followed by correspondence-specific preprocessing, manual ground-truth labeling, and structured dataset preparation. Model training employed intfloat/multilingual-e5-base with incremental N-shot experiments from 1-shot to 6-shot and contrastive representation learning using CosineSimilarityLoss. The 6-shot configuration, using six training documents per class, produced the best reported performance, achieving an accuracy of 85.87% and a weighted F1-score of 0.85. Training-loss and error-rate patterns also indicated stable model adaptation under the reported experimental conditions. The selected model was subsequently integrated into a lightweight Streamlit prototype. The findings demonstrate that SetFit provides an efficient approach for academic-document classification when institution-specific labeled datasets are scarce, while consistent label definition and real-time OCR integration remain important considerations for further system development. ]]>
Copyrights © 2026
