Prompt‐Aware Graph Reasoning and Explicit Ranking for Multi‐Trait Essay Scoring in Intelligent Education

ABSTRACT

Automated essay scoring (AES) plays a critical role in providing rapid and objective feedback in language assessment. However, cross‐prompt multi‐trait scoring faces two primary challenges: modelling complex topological dependencies between writing dimensions and overcoming the regression paradigm's fragility against feature shifts on unseen prompts, which often leads to blurred boundaries in fine‐grained. Current approaches frequently overlook these cognitive inter‐dependencies and fail to capture subtle cross‐essay discrepancies. To address these limitations, we propose Prompt‐Aware Graph Reasoning and Explicit Ranking (PAGER). Specifically, a decoupled feature extraction mechanism utilising a stop‐gradient strategy is first deployed to mitigate overfitting and prevent representation collapse. We then employ a prior‐guided sparse topology coupled with a dynamic semantic gating mechanism to adaptively regulate trait information routing across heterogeneous rubrics. Finally, to transcend the limitations of traditional scalar regression, PAGER leverages a Siamese architecture to embed explicit relative partial‐order learning within the graph latent space, reinforcing the model's discriminative resolution. Experimental results on the ASAP++ benchmark—spanning eight prompts across argumentative, narrative and source‐dependent genres with expert annotations on eight distinct writing traits—demonstrate that PAGER achieves new state‐of‐the‐art performance with an average quadratic weighted kappa (QWK) of 0.757, outperforming the strongest baseline by a significant relative margin of 17.91%, validating its enhanced cross‐prompt generalisability.