Leveraging Large Language Models to Detect Academic Anxiety in Indonesian English for Specific Purposes Students through Reflective Writing

Abstract

This study investigates the capacity of Large Language Models to identify academic anxiety in reflective writing produced by English for Specific Purposes students from Indonesia. It tackles two main issues: how well LLMs can identify anxiety from linguistic and environmental cues, and how anxiety-related language markers change depending on the type of activity and level of expertise. Employing a quantitative exploratory-correlational design, the study involved 600 undergraduate ESP students from Universitas Muhammadiyah Gresik. In addition to submitting two samples of reflective writing, each participant filled out a validated Academic Anxiety Inventory. To extract important language variables, such as lexical density, emotional Valence, modal usage, and syntactic complexity, transformer-based models (BERT, RoBERTa) were improved. Analytical reflections displayed greater lexical richness and syntactic complexity, but narrative reflections displayed more negative sentiment and hedging, according to MANOVA results, which demonstrated significant differences in anxiety markers. Higher-proficiency students demonstrated balanced rhetorical control and emotional tone, whereas lower-proficiency students exhibited greater signs of language anxiety. These results provide credence to the use of LLMs as non-invasive, scalable instruments for emotional diagnosis in ESP settings.