Lightweight Automated Short Answer Grading using Semantic Similarity

This preprint presents a lightweight automated short answer grading framework that combines Sentence-BERT (SBERT) semantic similarity with keyword-based scoring for evaluating CBSE-style short answers. The proposed system was evaluated on a dataset of 521 manually annotated responses and achieved a Pearson correlation coefficient of 0.76 while outperforming a TF-IDF baseline by approximately 35%. This work was conducted under the guidance of Prof. Niloy Ganguly, Department of Computer Science and Engineering, IIT Kharagpur. The manuscript has also been submitted to arXiv and is currently awaiting moderation.