Abstract: Prerequisite relation prediction for knowledge concepts seeks to enhance the curriculum knowledge graph by exploring semantic and topological dependencies among concepts,thereby improving downstream applications such as large-scale resource organization and personalized learning path planning.Existing methods,which mainly rely on feature engineering and deep lear-ning,still struggle to effectively model entity-level semantics and the directional nature of prerequisite relations,leaving room for further improvement.To address this problem,this paper proposes a direction-aware siamese network for knowledge concept prerequisite relation learning(DSN-PRL).Firstly,DSN-PRL employs a contrastive learning-based pre-trained language model,BERT,to capture fine-grained semantic representations of knowledge concepts.It then applies a graph neural network to incorporate multi-hop topological features and enhance hierarchical structure modeling.Finally,a direction-aware siamese network is designed to learn the directional distinctions between concept pairs for accurate prerequisite relation prediction.Experiments conducted on three benchmark datasets demonstrate that DSN-PRL outperforms existing baseline methods across multiple key evaluation metrics.In particular,compared with the best baseline model DGPL,DSN-PRL improves precision by 7.3 percentage points,2.7 percentage points,and 11.4 percentage points,and F1 by 1.6 percentage points,1.3 percentage points,and 4.3 percentage points,respectively.

Key words: Prerequisite relation prediction of knowledge concepts, Pre-trained language model, Contrastive learning, Graph neural network, Siamese network