Abstract
Automatic Question Generation (QG) in NLP research has garnered significant attention, aimed at assisting educators in improving classroom performance and promoting the adoption of intelligent education systems. However, few studies have addressed the critical challenge of controlling difficulty levels, particularly at the fine-grained, individual student level required for adaptive education. To achieve personalized question generation, we explored the integration of language models with knowledge tracing models, resulting in the development of the Language-Deep-Knowledge-Tracing (LDKT) model. This model can extract students’ textual interaction histories and represent their knowledge states, allowing for accurate predictions of their performance on new questions. Subsequently, the Difficulty-Controllable Question Generation (DCQG) model is guided to generate questions that align with the target difficulty, using datasets processed by the LDKT model. Both automated metrics and human evaluations have demonstrated the effectiveness of our proposed modules. The LDKT model achieved a maximum AUC of 0.77, while the DCQG model attained a minimum difficulty deviation RMSE of 3.96, both showing improvements over baselines. Further studies indicate that our work provides more options for modern intelligent education and advances the digitalization of teaching.