Automatic Short Answer Grading (ASAG) In Education: A Review Of AI And Deep Learning Techniques

Author: Ashvi Patel, Rupal Chaudhari, Mehul Patel

Published Online: January 5, 2026

Abstract

References

Automatic Short Answer Grading (ASAG) has become a vital research area in educational technology, aiming to provide scalable, efficient, and fair evaluation of student responses. Early studies using machine learning demonstrated the feasibility of reducing grading workload through handcrafted features and statistical classifiers. However, such approaches struggled with semantic variability and domain adaptation. The introduction of deep learning enabled richer semantic representation, improving grading accuracy and robustness across tasks. In recent years, transformer based models have become the dominant paradigm. BERT and its variants, including Sentence-BERT and hybrid extensions, have consistently outperformed traditional neural networks by capturing deep contextual embeddings and semantic similarity more effectively. Comparative studies further confirm BERT‘s superiority over earlier embedding based and RNN based approaches, while also revealing challenges related to interpretability and domain transfer. More recent explorations into large language models, such as GPT and T5, demonstrate strong zero-shot and few-shot capabilities, extending the potential of ASAG but raising concerns around transparency, fairness, and multilingual support.This review synthesizes findings across two decades of research, emphasizing the evolution from feature-driven methods to BERT centered deep learning approaches and recent advances with LLMs. Open challenges remain in dataset scarcity, interpretability, multilingual grading, and trustworthy deployment in real world classrooms. The paper concludes by outlining future directions that integrate hybrid deep learning LLM approaches, benchmark development, and ethical frameworks to advance reliable and equitable ASAG.

Keywords: Automatic Short Answer Grading (ASAG) . Machine Learning . Recurrent Neural Networks (RNNs) . Convolutional Neural Networks (CNNs) . Deep Learning . Transformer Models . Bidirectional Encoder Representations from Transformers(BERT) . Generative Pre-trained Transformer (GPT) . Text-to-Text Transfer Transformer (T5) . Large Language Models (LLMs) . Educational Assessment.

Download PDF Pages ( 170-176 ) Download Full Article (PDF)

←Previous Next →

SPU - Journal of Science, Technology and Management Research

An Official Publication of Sankalchand Patel University, Visnagar, Gujarat, India.

Volume - 2 | Issue - 2 (July-December 2025)

Automatic Short Answer Grading (ASAG) In Education: A Review Of AI And Deep Learning Techniques