Enhancing automated essay scoring with GCNs and multi-level features for robust multidimensional assessments

Abstract The advancement of automated essay scoring (AES) is pivotal in alleviating the burdens on educators and ensuring fair and reliable writing assessments. While deep neural networks have enhanced AES accuracy, they frequently encounter challenges in capturing comprehensive contextual features and achieving generalizability across multiple dimensions of writing quality. To address these limitations, we propose GCNs-MTL+Multi-Level Features, a novel integration of graph convolutional networks (GCNs) with the pretrained BERT model and multi-task learning (MTL). Our proposed approach significantly enriches essay representation fidelity by incorporating both word-level and sentence-level features, thereby enhancing transparency and improving the robustness and accuracy of writing evaluations across holistic and analytic rating scales. By sharing representations across AES tasks, GCNs-MTL+Multi-Level Features streamlines the evaluation process, establishing a new benchmark in multidimensional writing assessments.