Deep Learning Modeling of Teachers’ Reflective Texts Toward a Hybrid Workflow Integrating Transformer-Based Natural Language Processing and Reflexive Thematic Analysis

Abstract

Purpose: This study examines what teaching- and learning-related concerns teachers foreground in reflective texts about generative Artificial Intelligence (AI) in education. It further investigates how automated and human-coded approaches capture these concerns, aiming to develop a transferable workflow for large-scale yet context-sensitive educational text analysis. Theory: The analysis is framed by Discourse Analysis (DA), which conceptualizes language as social practice. Three analytic lenses, Positioning, Stance, and Metaphor/Framing, serve as criteria to compare automated and human-coded results and to interpret how teachers construct responsibility, identity, teaching, and learning in their talk about AI. Method: A parallel mixed-methods design was employed. On the automated side, Transformer-based Natural Language Processing (NLP) techniques were used to model thematic structures and sentiment patterns in teachers’ reflective texts. On the human side, Reflexive Thematic Analysis (RTA) was applied to workshop transcripts. The two sets of results were then systematically compared, with convergence and mismatch in identifying teaching- and learning-related themes forming the basis for a sequential human-machine workflow. Results: The analysis showed that teachers foregrounded five major teaching- and learning-related domains: teaching efficiency and support, individualized learning, ethics and governance, professional identity and responsibility, and risk and control. Both NLP and RTA captured these broad domains, but NLP modeling was more effective at identifying overarching topic structures and general sentiment trends, while it systematically misread conditional permission, irony, and metaphorical framing. RTA revealed these differences and reconstructed teachers’ boundary-setting and action scripts. Combining the two paths yielded a sequential workflow (“automated scanning - uncertainty alerts - human close reading - rule feedback - transparent auditing”), balancing efficiency with interpretive depth.