The evaluation of handwritten academic scripts has traditionally been a labor-intensive and subjective endeavor, susceptible to inconsistencies arising from evaluator fatigue, human bias, and high variability in handwriting styles. This paper presents a hybrid machine learning framework designed to transition from manual, time-intensive grading to a scalable, AI-driven automated assessment system. By integrating high-precision OCR pipelines (TrOCR) with deep semantic models such as BERT and Sentence-BERT (SBERT), the proposed system bridges the gap between physical script digitization and contextual understanding. The Janmitra modular architecture achieves a 70% reduction in evaluation workload and a peak semantic accuracy of 95.1% using SBERT-based vector embeddings and the Universal Sentence Encoder (USE). The framework incorporates a sentence-by-sentence comparison methodology providing granular pedagogical feedback, identifying missing conceptual points and flagging redundant content, establishing a robust foundation for objective, efficient, and fair academic assessment in digitally-transformed

The development of a unified framework integrating TrOCR and BERT represents a profound philosophical shift in academic assessment. By transitioning from the evaluation of “character strings” to the modelling of “knowledge states,” this system addresses the systemic crisis of manual grading. The objective metrics—notably the 97.6% QWK score—demonstrate that AI can achieve a level of consistency and fairness that human evaluators, constrained by fatigue and subjective bias, often cannot maintain at scale. The pedagogical value of this framework extends beyond mere scoring. By utilising the 2D similarity matrix and SHAP-based explainability, the system transforms assessment into a diagnostic tool. Students are no longer provided with a solitary grade but with a detailed map of their conceptual gaps. For the educator, the 70% reduction in evaluation time facilitates a transition from administrative grader to pedagogical mentor, allowing for more frequent, low-stakes continuous assessment that was previously logistically impossible. As educational institutions navigate the post-pandemic digital imperative, the inclusion of robust accessibility features ensures that the future of assessment is as equitable as it is efficient. The modularity of the Janmitra and ASSESS frameworks provides a scalable blueprint for global academic modernisation, ensuring that the subjective limitations of the past do not hinder the objective potential of future learners [1, 2, 3, 11].

[1]Mohanraj G et al., “An enhanced framework for smart automated evaluations of answer scripts using NLP and deep learning methods,” Multimedia Tools and Applications, 2023. [2]J. Clerk Maxwell, “Utilising BERT for Information Retrieval, Survey, Applications, Resources, and Challenges,” ACM Comput. Surv., Vol. 56, No. 7, 2024. [3]I. S. Jacobs and C. P. Bean, “AutoEval: A NLP Approach for Automatic Test Evaluation System,” IEEE 4th International Conference on Computing, Power and Communication Technologies (GUCON), 2021. [4]V. Agrawal, J. Jagtap, MVV Prasad Kantipudi, “An overview of Hand Drawn Diagram Recognition Methods,” IEEE ACCESS, 2024. [5]J. Kim, S. Park, A. Carriquiry, “A deep learning approach for handwritten document comparison using latent feature vectors,” The ASA Data Science Journal, 2024. [6]Y. S. Chernyshova, V. V. Arlazarov, and A. V. Sheshkus, “Two-Step CNN Framework for Text Line Recognition in Camera-Captured Images,” IEEE ACCESS, 2020. [7]A. Rokade et al., “Automated Grading System using Natural Language Processing,” 2nd International Conference on Inventive Communication and Computational Technologies, 2018. [8]D.R. Tetali et al., “A Python Tool for Evaluation of Subjective Answers (APTESA),” IJMET, Vol. 8, 2017. [9]M. Syamala Devi and H. Mittal, “Machine Learning techniques with Ontology for subjective answer evaluation,” IJNLC, Vol. 5, 2016. [10]C. Roy and C. Chaudhari, “Case-Based Modeling of answer points for semi-automated evaluation,” IEEE IACC, 2018. [11]K. Surya, E. Gayakwad, and Nallakaruppan M.K., “Deep learning for Short Answer Scoring,” IJRTE, Vol. 7, 2019. [12]D. Cer et al., “Universal Sentence Encoder,” 2018. [13]A. Graves et al., “A Novel Connectionist System for Unconstrained Handwriting Recognition,” IEEE TPAMI, 2009. [14]Y. LeCun et al., “Gradient-Based Learning Applied to Document Recognition,” Proceedings of the IEEE, Vol. 86, No. 11, 1998. [15]Alan Joseph et al., “Comparative Analysis of Text Classification Models for Offensive Language Detection,” IJERA, Vol. 04, 2024. [16]Anu Rose Joy, “An Overview of Fake News Detection using BiLSTM Models,” IJERA, Vol. 03, 2023. [17]Arun Robin et al., “Improved Handwritten Digit Recognition Using Deep Learning,” IJERA, Vol. 03, 2023. [18]N. Joseph and T. A. Thomas, “A Systematic Review of Content-Based Image Retrieval Techniques,” IJERA, Vol. 03, 2023. Era Johri et al., “ASSESS – Automated subjective answer evaluation using Semantic Learning,” K J Somaiya College of Engineering, 2021.

© 2025 International Journal of Engineering and Techniques (IJET).

Digital Object Identifier (DOI)

IJET assigns a unique DOI to every accepted article via Zenodo for persistent linking and citation. Your article’s DOI: https://doi.org/{{doi}}

Open DOI About Zenodo DOI

Close