Smart Method to Automatically Match Question and Answer in Multiple-Choice Test

Manually grading multiple-choice paper tests has numerous difficulties due to time-consuming and unintended errors. In this study, to minimize the problems, an automatic method of three stages was proposed to match questions and answers, which can be applied for grading. Its input is images of multiple-choice test papers on which a student answers by marking a circle around a letter, e.g. A, B, C, or D of a selected answer. From the input image, question ID, e.g. “Question 1” or “Câu 2” and the circle mark will be detected to create bounding boxes using a fine-tuned SAST text detection model. The second stage recognizes text in question ID and letters in circle mark bounding boxes using a lightweight fine-tuned text recognition model named SVTRLCNet. In the third stage, letter(s) in the circle mark(s) and corresponding question ID are matched. The stage includes 3 sub-steps namely (1) arranging question IDs and circle marks in the order from left to right and top to bottom based on the bounding box coordinates, (2) classifying question IDs and circle marks based on recognized text and letters, and (3) matching adjacent question ID and circle mark to make pairs of question ID and letter(s) in circle mark(s). AICS MultiChoice dataset, which includes mainly multiple-choice tests of math and English, was developed to fine-tune and test the method achieving 93.6% matching accuracy. In addition, the method was further evaluated on another self-developed test set, AICS MultiChoice Matching, which includes challenging scenarios, and achieved 91.2% matching accuracy.