Evaluating student answer scripts is a labor-intensive and subjective process, often influenced by human biases and inconsistencies. Automated Short Answer Grading (ASAG) seeks to address this challenge through computational methods that provide scalable and objective assessment. However, most existing ASAG approaches depend fully on supervised learning models that require extensive labeled datasets, which are costly to generate and require consensus among multiple evaluators. In this paper, a novel semi-supervised ASAG framework is proposed that significantly reduces the dependence on labeled data while maintaining grading accuracy. The proposed method employs the universal sentence encoder to generate powerful semantic embeddings of student answers and applies a Modified Density-Based Spatial Clustering algorithm to group similar responses. A reference-aware gradation mechanism is introduced for assigning scores to unlabeled responses based on cluster properties and their similarity to the reference answer. The model is validated on multiple benchmark datasets such as ASAP, Texas, SciEntsBank, and a newly curated dataset from Indian university students. The method provides 0.66 and 0.975 as Pearson correlation coefficient and Root Mean Square Error for the ASAP dataset, respectively. Similarly, for SciEntsBank, Texas, and CS2 datasets, these values are (0.60, 0.916), (0.67, 1.053), and (0.64, 1.161), respectively. Although the results are not significantly improved over the recent deep learning-based models but they are comparable and are achieved with less computational cost with a small amount of labeled data. The experimental results demonstrate that the proposed semi-supervised model presents a competitive performance against the state-of-the-art deep learning-based supervised approaches, even using a small fraction of labeled data for training. These findings establish the effectiveness and applicability of the proposed semi-supervised method for automated short answer grading in real-world educational environments.
Haldar et al. (Mon,) studied this question.