Automatic Subjective Answer Evaluation

Abstract: Automatic evaluation of subjective answers has become a vital area of research due to its potential to reduce the

manual effort required in educational assessments. This paper presents an advanced system for the automatic evaluation of

handwritten subjective answers, integrating Optical Character Recognition (OCR), Natural Language Processing (NLP), and

semantic similarity techniques. The system employs the Google Cloud Vision API to extract textual data from handwritten

answer sheets with high accuracy. Extracted responses undergo preprocessing, including spell correction, and are semantically

compared with ideal answers using both BERT and fine-tuned SBERT models. To enhance grading reliability, a custom

contrastive learning mechanism is implemented for SBERT fine-tuning, using student-ideal answer pairs. The evaluation is

performed via a Flask-based backend, which also supports training workflows through API endpoints. Feedback and marks are

generated based on semantic similarity and model confidence. This system demonstrates an effective solution for automating

subjective answer assessment with a high degree of flexibility and accuracy, particularly for handwritten inputs. The system

supports both real-time evaluation and model customization, offering educators the flexibility to retrain models using domainspecific datasets. A user-friendly web interface allows for seamless uploading of answer images, configuration of model settings,

and visualization of results. Additionally, the system integrates a secure user authentication module for access control, enabling

personalized model training and usage history tracking. Experimental results demonstrate that the fine-tuned SBERT model

significantly improves semantic alignment with ground-truth answers, especially in the context of varied handwriting styles and

non-standard grammar.