A Rubric-Centric Approach for Automated Test Correction Utilizing RAG and Fine Tuning

Assessments are a popular method to quantify a person's knowledge, skill, or cognitive ability. However, to grade such assessments, particularly subjective questions requires a significant amount of time and effort for the examiners. There is also a degree of variance due to inconsistency from different examiners. Thus institutions often resort to taking objective-type questions, which may not completely capture a person's understanding. Unfortunately, many other proposed solutions, tend to be difficult to implement and scale. To address these issues, this paper proposes an innovative method to automate the grading of subjective assessments by leveraging recent advances in artificial intelligence. This approach aims to replicate the grading style of human experts, offering significant benefits for educational institutions, faculty, and students by enhancing the efficiency and quality of grading. Widespread adoption can help increase the quality of education overall, resulting in more qualified graduates. This method involves using a rubric to fine- tune pre-trained language models, providing them with clear guidelines to ensure consistent and accurate grading. Inspired by how rubrics guide human examiners to minimize variation, this approach will use rubrics to instruct the model on the most relevant grading criteria. Utilizing such clearly defined and precise grading criteria, users can obtain reliable results, comparable to human experts.