Classifying Mathematics Teacher Questions to Support Mathematical Discourse.

This paper examines whether natural language processing technologies can provide teachers with high-quality formative feedback about questioning practices that promote rich inclusive mathematical discourse within classrooms. This paper describes how a training dataset was collected and labeled using teacher questioning classifications that are grounded in the mathematics education literature, and it compares the performance of four classifier models fine-tuned using that dataset. Of the models tested, we find that RoBERTa, an open-source LLM, had a 76% accuracy in classifying questions. These modern transfer-learning based approaches require significantly fewer data points than traditional machine-learning methods and are ideal in low-resource scenarios like question classification. The paper concludes by discussing potential use cases within the field of mathematics teacher education and describes how the classifier models created can be publicly accessed.