Exploring Fairness in Automated Grading and Feedback Generation of Open-Response Math Problems

The rapid growth and development of NLP techniques have resulted in Computer-Based Learning Platforms (CBLPs) leveraging innovative approaches toward automated grading and feedback generation of open-ended problems. Researchers have explored these techniques in driving a varying range of interventions that range from assessing the quality of the work and recommending changes to the answers that can enhance the quality of the responses for students to automated grading and feedback generation of responses for teachers. A crucial aspect of the automated assessment of student response is identifying and addressing fairness and equity issues in an educational context, as academic performance can impact the types of opportunities available to the students. While prior works have conducted posthoc analysis exploring aspects of algorithmic fairness of various models, the assessment of open-ended answers is often subjective. Teachers leverage contextual knowledge such as the perception of the student effort or students' prior knowledge. While such factors exist, it is not obvious how data from the teacher can introduce biases or introduce measurable risks to the fairness and equity of the NLP models. In this paper, we build on our prior analysis of the grading behavior of teachers on open-ended math problems for middle school students and explore possible next steps we can take to expand on our work. First, we propose a simulation study to explore the various risks associated with Human-AI interaction in the automated grading of open-ended problems. Second, we propose an extensive study expanding on our work to generate grades for open responses when a student is anonymized vs. not anonymized.