A knowledge graph-enhanced adaptive recommendation framework for college english vocabulary learning with learner state modeling

Vocabulary recommendation plays an important role in the successful acquisition of a second language, but the existing method is not able to balance semantic coherence with the language learners’ learning trajectories. We present a framework that learns vocabulary embeddings and learner representations simultaneously based on a knowledge graph structure with temporal learning dynamics. We build a multi-relational knowledge graph describing the semantic, morphological, and prerequisite relations between the vocabulary items and use graph convolutional networks to propagate structural information across the graph while modeling the knowledge state in the learner model by the exponential forgetting. Our approach is different from previous approaches with fixed semantic representations, since a vocabulary embedding can be changed according to patterns of learning difficulties between learners, and the learner’s representation will change over a space of semantics that captures pedagogically relevant geometry. We introduce a semantic coherence constraint which ensures that recommended vocabulary are grouped together into coherent clusters, facilitating association learning. Results for the actual learning gain in 487 university students learning 1856 English vocabulary items show 5.1 to 5.5% gains for rankings metrics over state-of-the-art baselines, and 12.2% gains over state-of-the-art baselines for actual learning gain. The learned embedding space has pedagogical emergent structure, with basic vocabulary in central hub regions and advanced vocabulary in peripheral specialized regions. The ablation experiments reveal that the joint training brings 19.1% cumulative gain compared to static embeddings while learnable representations and graph convolution methods give the highest value gains individually.