Programming knowledge tracing is a task aimed at assessing students’ mastery of knowledge concepts during their programming exercise process. Existing programming knowledge tracing models typically overlook the impact of semantic information in programming question texts on students’ answering performance. Additionally, most models are based on Recurrent Neural Networks (RNNs), which have defect with long-range temporal dependencies and vanishing gradient problem. To address these deficiencies, we propose a novel model (GTPKT) that leverages Generative Pre-trained Transformer (GPT) on question texts to extract a contextual representation of the question, effectively capturing semantic information and similarity relationships. Subsequently, we integrate the generated question embedding with students’ interaction data to enrich the model’s input layer. Finally, we used Temporal Convolutional Network (TCN) to model students’ knowledge states and predict their future programming performance. Extensive experiments on two publicly available datasets represent that our GTPKT achieves superior performance in prediction.