Optimizing the Online Learners' Verbal Intention Classification Efficiency Based on the Multi-Head Attention Mechanism Algorithm

Abstract

To analyse speech intention based on discussion texts in online collaborative discussions, automatic classification of discussion texts is conducted to assist teachers improve their abilities to diagnose and analyse the discussion process. The current study proposes a deep learning network model that incorporates multi-head attention mechanism with bidirectional long short-term memory (MA-BiLSTM). The proposed algorithm acquires contextual semantic connections from a global perspective and the role of key feature words within sentences from a local perspective to further strengthen the semantic features of the texts. The proposed model was employed to classify 12,000 interactive texts generated during online collaborative discussion activities. Results show that MA-BiLSTM achieved an overall classification accuracy of 83.25%, which is at least 2.83% higher than those of other baseline models. However, the classification of consultative and administrative interactive texts is minimally effective. MA-BiLSTM achieved better than the existing classification methods for interactive text classification.