Semi-Supervised Bullying Detection in Narrative Student Counselling Reports Using a Hybrid CNN-LSTM with Pseudo-Labelling

Abstract

Bullying incidents in schools are often documented in narrative student counselling reports containing informal language, emotional expressions, and contextual dependencies, which pose challenges for automated text classification, particularly under limited labeled data conditions. This study aims to develop a bullying detection model for narrative student counselling reports using a Hybrid CNN-LSTM architecture combined with a pseudo-labelling-based semi-supervised learning approach. The proposed model is trained through a two-stage process, consisting of pre-training on approximately 70,000 publicly available abusive-language texts and fine-tuning using 1,000 anonymized student counselling reports validated by guidance counsellors. Pseudo-labelling is employed to expand the training data while preserving domain relevance and adhering to ethical considerations. Experimental results show that the proposed model achieves an accuracy of 0.8698, a recall of 0.8570, and an F1-score of 0.7951. Although the precision value (0.7415) is relatively lower, higher recall is prioritized to reduce the risk of overlooking potential bullying cases in the school counselling context. Comparative analysis with Logistic Regression and Linear SVM indicates that the Hybrid CNN-LSTM model demonstrates more stable performance when processing longer narrative inputs that require contextual interpretation. This study contributes empirical evidence on the effectiveness of semi-supervised deep learning for bullying detection in low-resource, narrative student counselling data, a setting that remains underexplored in prior work.