Improving Automatic Short Answer Scoring Task Through a Hybrid Deep Learning Framework

Abstract

An automatic short-answer scoring system involves using computational techniques to automatically evaluate and score student answers based on a given question and desired answer. The increasing reliance on automated systems for assessing student responses has highlighted the need for accurate and reliable short-answer scoring mechanisms. This research aims to improve the understanding and evaluation of student answers by developing an advanced automatic scoring system. While previous studies have explored various methodologies, many fail to capture the full complexity of response text. To address this gap, our study combines the strengths of classical neural networks with the capabilities of large language models. Specifically, we fine-tune the Bidirectional Encoder Representations from Transformers (BERT) model and integrate it with a recurrent neural network to enhance the depth of text comprehension. We evaluate our approach on the widely-used Mohler dataset and benchmark its performance against several baseline models using RMSE (Root Mean Square Error) and Pearson correlation metrics. The experimental results demonstrate that our method outperforms most existing systems, providing a more robust solution for automatic short-answer scoring.