A generative AI-driven intelligent diagnosis system for teaching quality: multi-source fusion of classroom discourse and temporal behavior patterns via NLP

This study introduces a multimodal intelligent diagnosis system for assessing teaching quality in higher education, integrating natural language processing (NLP) of classroom discourse with temporal sequence modeling for teaching behavior pattern recognition. Leveraging 86 classroom sessions from the MICASE corpus, the system utilizes a fine-tuned BERT model for discourse function classification, a BiLSTM-CRF model for temporal teaching behavior recognition, and a LoRA fine-tuned LLaMA-2-7B-Chat model for automated diagnostic report generation. Experimental results demonstrate the effectiveness of the system, achieving a macro-F1 of 79.41% for discourse classification and 74.60% for behavior pattern recognition. Expert evaluation indicates moderate-to-good accuracy (3.54/5), practicality (3.31/5), and readability (3.92/5), suggesting the reports are usable as a supplementary, formative aid rather than a stand-alone judgment. The system fuses two text-derived information streams—discourse-function distributions and temporal behavior patterns—to support automated teaching quality evaluation. The present implementation is text-based; integration of audio and visual modalities is identified as future work.