Intelligent english education resource recommendation via LLM-knowledge graph integration and ecological niche...

Teaching English to large and varied student populations has made adaptive, learner-aware resource recommendation a pressing need, yet most deployed systems still lean on collaborative filtering or content heuristics that read learners only at the surface. Such systems tend to stumble in four recurring ways: they miss the finer grain of linguistic competence, they reason over sparse and brittle knowledge associations, they falter whenever a learner or a resource is new, and they rarely travel well to settings their training never saw. We address these problems together through a framework in which a large language model and a knowledge graph inform one another, their match to each learner mediated by a fitness function adapted from ecological niche theory. The language model reads a learner's own writing to estimate evolving competence and proposes relations that fill thin regions of the graph, while the graph keeps the model's inferences anchored to the prerequisite and difficulty structure of English. A niche fitness score then judges how closely each resource's demands meet a learner's needs, and a path generator orders the survivors into prerequisite-respecting study sequences. On logs from 3,214 learners, 12,486 resources, and 287,653 interactions, the method reaches a Precision@5 of 0.782 and an NDCG@10 of 0.724, ahead of five baselines. Ablation was, frankly, the part that surprised us: the niche filter matters more to ranking quality than the model's semantic re-scoring. A three-axis generalization protocol, splitting transfer across learners, resource types, and difficulty levels, yields a composite index of 0.863, with cold-start decay held to 18.3%.