An Exploratory Study on YOLO-v8-Based Automatic Scoring for Korean Sign Language SignWriting Tasks by L1 (Korean) Learners: Focusing on Visual Feedback

This exploratory study proposes a YOLO-v8-based automatic scoring procedure to reduce instructors’ grading workload and provide immediate feedback for Korean Sign Language SignWriting tasks performed by L1 (Korean) learners. A dataset of 25,000 images was constructed using 50 target classes—33 single-word signs, 16 compound expressions, phrases, or short sentences, and one long sentence—with 500 images per class. These were randomly synthesized without duplication into 13-page assignments (100 copies each; 1,300 pages in total), formatted in a five-line repetition layout. The trained YOLO-v8m model demonstrated robust detection performance ([email protected] = 0.995, precision = 0.999, recall = 1.000). Confidence scores, significantly correlated with human ratings, were visualized using color-coded bounding boxes and averaged scores, supporting immediate corrective feedback while reducing grading time by over 99.7%. In error simulation tests, the model delivered accurate visual feedback at the word level. However, for the long sentence task, the IoU dropped to 0.697, suggesting a need for layout and scale adjustments to enhance recognition stability. This study is the first to validate the feasibility of automatic scoring in Korean Sign Language SignWriting education. Future research will aim to enhance the precision of automatic scoring and verify changes in learner achievement and the effectiveness of instructional design through error-type analysis.