Automated essay scoring (AES) systems typically rely on either sequential text models for contextual semantics or graph-based approaches for discourse structure, but rarely integrate both effectively. We present GDBERT-Score, a hybrid architecture that combines DeBERTa’s disentangled attention with graph convolutional networks (GCN) for holistic essay scoring. Each essay is represented as a sentence-level semantic graph, where nodes are DeBERTa mean-pooled sentence embeddings and edges connect sentence pairs whose cosine similarity exceeds a learned threshold ( τ = 0 . 4 ). GCN-derived structural embeddings are concatenated with DeBERTa’s document-level contextual embedding and fused through a progressive reduction network ( 896 → 512 → 256 → 128 ) to produce a final holistic score. Evaluated on the Kaggle Automated Essay Scoring 2.0 benchmark via six-fold cross-validation, GDBERT-Score achieves QWK = 0.7777 ± 0.0017 (mean ± std across 5 random seeds), significantly outperforming both the DeBERTa-only baseline ( p < 0 . 0001 ) and the TF–IDF graph variant ( p = 0 . 0010 ). Ablation experiments reveal that node feature quality, rather than graph topology, is the primary determinant of GCN effectiveness in automated essay scoring: replacing surface-level node features with DeBERTa sentence embeddings yields significant improvements over both the DeBERTa-only baseline ( Δ QWK = +0.0294, p < 0 . 0001 ) and the graph-augmented variant ( Δ QWK = +0.0239, p = 0 . 0010 ), confirmed across five random seeds (QWK = 0.7777 ± 0.0017).
Mi et al. (2026) studied this question.