Mining Improvement Points for University Teaching Feedback Based on BERT Embedding and Sentiment Convolutional Network

Abstract

University teaching feedback contains contextual semantics, implicit sentiment, and fragmented improvement suggestions, which makes automatic extraction difficult when only keyword or shallow syntactic features are used. To improve semantic understanding and emotion-aware localization, this study proposes a teaching-feedback mining model combining BERT embedding, a multi-channel sentiment convolutional network, emotional attention, and BiLSTM-CRF sequence labeling. After text cleaning, sentence segmentation, stop-word removal, and BIO-based sentiment annotation, BERT is used to generate deep contextual representations. Multi-scale convolutional kernels then extract local sentiment patterns, and an emotional attention mechanism enhances features highly related to improvement-point identification. The fused semantic and sentiment features are jointly modeled through a dual-task structure for sentiment classification and sequence labeling. Experiments on 18,000 teaching-feedback samples show that the model achieves an accuracy of 87.2%, an F1 score of 84.1%, and an improvement-point completeness of 80.5%, outperforming static word embeddings, LSTM embeddings, and single-task baselines. The proposed method provides an engineering-oriented text-signal processing framework for semantic feature extraction, sentiment modeling, and intelligent feedback-analysis systems.