Enhancing neural automatic short answer grading with attention masking, WordNet-based semantic enrichment, and ordinal regression

Automatic Short Answer Grading (ASAG) is an important task in educational technology and natural language processing (NLP), enabling efficient and consistent evaluation of open-ended student responses. Although recent neural architectures have achieved strong performance, further improvements are possible in model focus, semantic understanding, and score consistency. This study builds upon the hybrid convolutional neural network–bidirectional long short-term memory (CNN-BiLSTM) Siamese model by Chen et al. (2023) and introduces three complementary extensions. First, an attention masking mechanism is applied to enhance focus on relevant tokens and reduce noise from padding or less informative text. Second, semantic enrichment is achieved through the integration of WordNet-based lexical similarity, enabling the model to capture deeper relationships between conceptually related terms. Third, rank-consistent ordinal regression replaces categorical prediction to better reflect the ordered nature of grading. Experiments on three datasets (Haidir, Mohler, and Rahutomo) demonstrate that each enhancement improves performance, with the combined model achieving the strongest overall results. On average across datasets, the proposed framework reduces normalized mean absolute error (NMAE) from 0.117 to 0.087 and normalized root mean squared error (NRMSE) from 0.205 to 0.159, while increasing Pearson correlation from 0.703 to 0.806 and Spearman correlation from 0.708 to 0.792. These findings show that integrating attention refinement, external semantic knowledge, and ordinal modelling substantially enhances ASAG accuracy and ranking reliability. This work contributes to the development of more robust, linguistically grounded, and educationally meaningful ASAG systems for diverse learning contexts.