<![CDATA[The rapid transition to online learning during the COVID-19 pandemic has accelerated the demand for automated assessment systems capable of evaluating students’ understanding objectively, consistently, and efficiently. Automatic Short Answer Scoring (ASAS) has emerged as a central challenge within Natural Language Processing (NLP), particularly for assessing short or open-ended answers. This research proposes an ASAS framework that integrates Reference Answer Generation using Transformer-based models (BART and GPT-2) with the Maximum Marginal Relevance (MMR) method. While MMR has been widely used as an extractive technique to select diverse student answers as reference candidates, its performance remains limited by the availability and variability of student responses. To address this limitation, the proposed approach supplements MMR-generated references with paraphrased reference answers generated by fine-tuned Transformer models, thereby increasing lexical and syntactic diversity. Experimental evaluation was conducted on the Texas Short Answer Corpus, consisting of 2,442 student responses across 10 tasks. The results demonstrate that the proposed hybrid approach improves scoring performance, as indicated by an increase in Pearson correlation and a reduction in RMSE and MAE compared to the baseline MMR-based method.  ]]>
Copyrights © 2026
