In this paper, a knowledge graph construction scheme integrating deep learning and lightweight architecture is proposed for intelligent sports teaching scenarios, focusing on solving the three major problems of fuzzy entity recognition, computational redundancy, and low efficiency of knowledge fusion in the sports domain. We design a BERT-BiLSTM-CRF entity recognition model enhanced by attention mechanism, and combine it with TF-IDF weighting + alias dictionary matching strategy to improve the entity linking accuracy. The DeLighT module is then introduced to optimize the Transformer, and the parameter distribution is dynamically adjusted through the expansion-scaling mechanism, which significantly reduces the computational redundancy while maintaining the performance. Finally, based on the ontology “seven-step approach” to build the knowledge system of physical education courses, using Neo4j graph database to realize the efficient storage of ternary groups. The BERTBiLSTM-CRF head entity detection model has an F1 value of 91.03%, with a compressed number of model parameters after optimization by the DeLighT module, and a composite score of 92.13%, which is an improvement of 14.19 points from the baseline. Teaching empirical evidence shows that the system significantly improves the training effect, the experimental group of students’ physical health indicators are better than the control group in all aspects, lung capacity is improved by 12.1% (3,142.75ml and 2,803.64ml), the 50-meter run is accelerated by 0.84 seconds (8.01s and 8.85s), and the standing long jump is increased by 11.2% (211.95cm and 190.72cm). In the dimension of sports learning interest, positivity increased by 45.5% (4.73±0.59 and 3.25±1.25) and negativity decreased by 71.5% (1.09±0.35 and 3.82±1.06), which verified the effectiveness of knowledge graph-driven intelligent teaching in personalized training instruction and learning motivation.
