Efficient automated evaluation of answer scripts using LLMS, NLP, and deep learning
S. B. Ashoka, K. Sai Deep, Gagan Goutham, Syed Mohammad Muzammil Ganihar, P K Udayaprasad, G C Lakshmikantha, P. Dayananda · 2025
NLP and DL researchers are increasingly concentrating on automated answer script evaluation as an attempt to overcome the limitations and inconsistencies of traditional grading. semantic parallelism and contextual preserving are problems for conventional techniques like WordNet, OCR, and Word2Vec. For enhanced accuracy and efficiency, this research presents a novel hybrid architecture that integrates advanced DL and NLP techniques. The system uses a Cosine similarity network for accurate similarity scoring, an OCR model for handwritten text transformation, Universal Sentence Encoder for embeddings, and LLM for contextual analysis. A Deep Columnar CNN additionally handles complex answer formats and symbols. Considering 93.8% accuracy, 94.1% precision, 92.7% recall, and a 93.4% F1-score, the integrated technique significantly improves conventional techniques in terms of criteria for evaluation. Because of its cloud-based architecture, that guarantees scalability and real-time processing, automated grading in educational environments can be effectively implemented.