A BERT-Based Model for Determining University Students' Non-Clinical Academic Difficulties

Abstract

Academic issues like procrastination, poor time management, high stress, and low mindfulness often compromise university students' academic outcomes. These concerns, though non-clinical, can profoundly influence academic achievement and personal growth. This work presents a BERT-based classifier specially trained to detect such non-clinical academic issues via student-generated text data. A synthetic dataset comprising 50,000 utterances was created to mimic real-world student concerns, enabling intense model training while maintaining privacy and ethical requirements. Although a small set of anonymized student responses was collected for qualitative reference, it was not used in training or evaluation due to privacy and data volume limitations. With state-of-the art training strategies, including gradient checkpointing, class weighting, and automatic mixed precision, the fine-tuned BERT classifier attained a validation accuracy of 92.3% and F1-score of 0.91. This outcome showcases BERT's ability to differentiate subtle psychological states in non-clinical contexts. The suggested approach provides a scalable, privacy-friendly, and explainable basis for personalized academic support platforms, especially in university settings where students avoid seeking professional assistance. This work contributes to AI-driven educational technology and facilitates the development of adaptive, non-clinical student well-being interventions in the future.

Description

Citation

Wickramaarachchi, W.K.R.D.; Buddhika, R.A.P. (2025) A BERT-Based Model for Determining University Students' Non-Clinical Academic Difficulties, Technology Innovations & Digitalization Accompanying Creativity (TIDAC), NSBM Green University, Sri Lanka. P.330

Endorsement

Review

Supplemented By

Referenced By