Scholarlink

<p>In recent years, academic competitions at various levels and across numerous fields have emerged rapidly, with student participation continually increasing. As the number of participating teams grows, the workload for reviewing papers has increased sharply, resulting in greater demands on human resources. This has created an objective need for initial machine scoring followed by secondary human evaluation. Automated Essay Scoring (AES) systems can significantly reduce the workload and improve the fairness of paper evaluations. Consequently, research that focuses on improving AES performance has increased substantially. However, current AES systems either heavily depend on expert-defined rules or lack interpretability. Ensuring flexibility and transparency in scoring standards remains a challenge in the AES field. In this study, we explore an automated scoring method that combines expert models with deep learning and propose an essay scoring framework, EMBR-AES. This framework extracts shallow features based on universally applicable expert models, captures deep semantic features using the Albert model, and then uses a neural network to identify latent patterns and structures associated with scoring. We conducted model experiments based on a small competition dataset. Results indicate that this method demonstrates good adaptability across different datasets and better reflects the true scoring standards. In terms of efficiency, EMBR-AES significantly reduces the scoring time, with the review time per paper limited to 25&#xa0;s, which is sufficient for most competitions.</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

An automated essay scoring approach based on expert models and deep learning

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • Runze Zhang,
  • Yang Zhang,
  • Yating Zhao,
  • Bin Jia,
  • Wenjuan Lian

摘要

In recent years, academic competitions at various levels and across numerous fields have emerged rapidly, with student participation continually increasing. As the number of participating teams grows, the workload for reviewing papers has increased sharply, resulting in greater demands on human resources. This has created an objective need for initial machine scoring followed by secondary human evaluation. Automated Essay Scoring (AES) systems can significantly reduce the workload and improve the fairness of paper evaluations. Consequently, research that focuses on improving AES performance has increased substantially. However, current AES systems either heavily depend on expert-defined rules or lack interpretability. Ensuring flexibility and transparency in scoring standards remains a challenge in the AES field. In this study, we explore an automated scoring method that combines expert models with deep learning and propose an essay scoring framework, EMBR-AES. This framework extracts shallow features based on universally applicable expert models, captures deep semantic features using the Albert model, and then uses a neural network to identify latent patterns and structures associated with scoring. We conducted model experiments based on a small competition dataset. Results indicate that this method demonstrates good adaptability across different datasets and better reflects the true scoring standards. In terms of efficiency, EMBR-AES significantly reduces the scoring time, with the review time per paper limited to 25 s, which is sufficient for most competitions.