Collaborative Project: Common Error Diagnosis and Support in Short-answer Math Questions

Project description

One important way to help struggling students improve in math is to deliver personalized support that addresses their specific weaknesses. Many math questions have common wrong answers (CWAs) that correspond to specific errors students make during their answering process, caused by misconceptions or a general lack of knowledge on certain math skills. To date, CWA identification and support remains a labor-intensive process at a limited scale because it requires significant effort by teachers and/or domain experts. In this project, the investigators will develop artificial intelligence (AI)-based mechanisms that can automatically identify CWAs from students’ answers to short-answer math questions and diagnose errors. Once these errors are identified, the investigators will enlist the help of teachers to design feedback and support mechanisms in various formats such as textual feedback messages and short videos. In turn, the investigators will integrate these diagnosis and effective support mechanisms into a teacher interface to support them in either classrooms or online learning environments. Overall, this project has the potential to lead to i) better understanding of CWAs in math questions and the underlying errors and ii) effective CWA support mechanisms for each error type. The project will be grounded in ASSISTments, a free web-based learning platform, therefore directly benefiting the 500,000 US students and 20,000 teachers using it and potentially an even larger number of students and teachers through the dissemination of research findings.
This project consists of four main research activities. First, the investigators will leverage math expression embedding methods to learn the representations of student errors by clustering CWAs across multiple questions in the latent math expression embedding vector space. These learned representations will enable the automated diagnosis of student errors in real time. Second, the investigators will develop new knowledge tracing algorithms that go beyond typical correctness analysis and analyze the full answer each student submits to each question. These algorithms will enable the automated tracking of students’ progress in correcting their errors. Third, the investigators will crowdsource multiple types of student support from teachers and integrate both student error diagnostics and support mechanisms into the existing ASSISTments teacher interface. This interface will provide feedback to teachers on which students are struggling in real time and recommend a support, which the teacher can either adopt and customize or reject and create their own support instead. Fourth, the investigators will conduct a randomized controlled trial to evaluate the effectiveness of each support mechanism in helping students correct their errors. This experiment will identify which support mechanisms are most effective at helping students correct each error type and improving learning outcomes.

Publications related to this project

M. Zhang, Z. Wang, R. G. Baraniuk, and A. S. Lan, "Math Operation Embeddings for Open-ended Solution Analysis and Feedback," International Conference on Educational Data Mining (EDM), June 2021

M. Zhang, S. Baral, N. Heffernan, and A. S. Lan, "Automatic Short Math Answer Grading via In-context Meta-learning," International Conference on Educational Data Mining (EDM), July 2022
Z. Wang, M. Zhang, R. G. Baraniuk, and A. S. Lan, "Scientific Formula Retrieval via Tree Embeddings," IEEE International Conference on Big Data, Dec. 2021

Z. Wang, R. G. Baraniuk, and A. S. Lan, "Math Word Problem Generation with Mathematical Consistency and Problem Context Constraints," Conference on Empirical Methods in Natural Language Processing (EMNLP), Nov. 2021

Back to my research page