Semantic analysis and intervention generation for dialogue between teachers and young children using large language models

While artificial intelligence holds immense promise for early childhood education, its deployment for young children faces significant challenges due to their unique communication patterns. Children typically exhibit fragmented logic and intense emotional expression, and standard language models respond with adult-like formulations that risk exceeding limited working memory and provide little cognitive scaffolding. To address this problem, this paper introduces a neuro-symbolic Bayesian graph transformer framework. This methodology integrates structural graph networks with probabilistic inference to robustly decode ambiguous expressions. Crucially, the framework mathematically incorporates cognitive load theory into the optimization objective, explicitly penalizing the generation of excessively complex vocabulary. On an age-adapted corpus reconstructed from authentic teacher–learner dialogue, the framework reaches 91.82% accuracy in intent recognition, a 7.22%-point improvement over the strongest comparison system. Furthermore, the model-estimated extraneous load of generated interventions is reduced by 1.45 units relative to the strongest baseline, and expert teachers rated its outputs as more pedagogically appropriate.