Scholarlink

<p>Automated essay scoring (AES) has evolved from accuracy-centric systems toward models emphasizing interpretability and learner-centered support. To address the growing demand for personalized feedback, we introduce, for the first time in this domain, a hybrid CNN–Transformer architecture within the Trait-Aware and Student-Adaptive Feedback (TAS-AF) framework. TAS-AF jointly models trait-specific, prompt-aware, and learner-conditioned features in a unified pipeline. The CNN–Transformer backbone captures both local syntactic patterns and global semantic dependencies under a multi-task learning paradigm, supporting fine-grained trait scoring and interpretable feedback generation. Feedback is further guided by multi-task trait embeddings, prompt contextualization, and essay-level proxy learner representations, while a trait-weighted aggregation mechanism produces both detailed, rubric-aligned feedback and concise holistic summaries. Experiments on the ASAP dataset demonstrate that TAS-AF achieves the highest average QWK of 0.6244, outperforming all state-of-the-art baselines. These results highlight that the integration of the CNN–Transformer architecture with trait- and learner-aware mechanisms not only enhances predictive accuracy but also produces feedback that is pedagogically meaningful and interpretable. Limitations include reliance on sufficient learner representation and increased computational overhead. Future work will explore few-shot learner modeling, lightweight adapter modules for scalability, and extensions to multilingual and cross-domain writing tasks.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CNN–transformer-based trait-aware scoring and student-aware feedback generation for English writing

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月28日

查看全文
  • Fang Lin

摘要

Automated essay scoring (AES) has evolved from accuracy-centric systems toward models emphasizing interpretability and learner-centered support. To address the growing demand for personalized feedback, we introduce, for the first time in this domain, a hybrid CNN–Transformer architecture within the Trait-Aware and Student-Adaptive Feedback (TAS-AF) framework. TAS-AF jointly models trait-specific, prompt-aware, and learner-conditioned features in a unified pipeline. The CNN–Transformer backbone captures both local syntactic patterns and global semantic dependencies under a multi-task learning paradigm, supporting fine-grained trait scoring and interpretable feedback generation. Feedback is further guided by multi-task trait embeddings, prompt contextualization, and essay-level proxy learner representations, while a trait-weighted aggregation mechanism produces both detailed, rubric-aligned feedback and concise holistic summaries. Experiments on the ASAP dataset demonstrate that TAS-AF achieves the highest average QWK of 0.6244, outperforming all state-of-the-art baselines. These results highlight that the integration of the CNN–Transformer architecture with trait- and learner-aware mechanisms not only enhances predictive accuracy but also produces feedback that is pedagogically meaningful and interpretable. Limitations include reliance on sufficient learner representation and increased computational overhead. Future work will explore few-shot learner modeling, lightweight adapter modules for scalability, and extensions to multilingual and cross-domain writing tasks.