Scholarlink

Automated Essay Scoring (AES) often suffers from sparse and imbalanced training data. We present an in-context learning (ICL) data-augmentation framework that prompts DeepSeek-V3 with six score-specific exemplars to synthesise 1200 additional essays. Training a CNN–LSTM–attention scorer on the augmented ASAP Prompt-1 set lifts Quadratic Weighted Kappa from 0.75 to 0.82, outperforming Easy Data Augmentation and NLPaug baselines. Kolmogorov–Smirnov and Mann–Whitney tests ( \(\alpha {=}0.05\) ) show no significant differences between original and generated texts in paragraph length, sentence complexity, topic relevance and grammatical error rate. Similar gains (+0.04 QWK) are observed on the LAB-AES2 benchmark. The method is zero-shot, model-agnostic and lightweight, offering a practical recipe for high-fidelity augmentationunder data scarcity. Our implementation is available at https://github.com/cactusYuri/llmbased_ASAP_AUG_main .

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

Data Augmentation for Automated Essay Scoring Using Large Language Models and In-Context Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yuli Zhang,
  • Yihan Luo,
  • Shenglu Hua,
  • Yuanlong Yu,
  • Bokai Cui

摘要

Automated Essay Scoring (AES) often suffers from sparse and imbalanced training data. We present an in-context learning (ICL) data-augmentation framework that prompts DeepSeek-V3 with six score-specific exemplars to synthesise 1200 additional essays. Training a CNN–LSTM–attention scorer on the augmented ASAP Prompt-1 set lifts Quadratic Weighted Kappa from 0.75 to 0.82, outperforming Easy Data Augmentation and NLPaug baselines. Kolmogorov–Smirnov and Mann–Whitney tests ( \(\alpha {=}0.05\) ) show no significant differences between original and generated texts in paragraph length, sentence complexity, topic relevance and grammatical error rate. Similar gains (+0.04 QWK) are observed on the LAB-AES2 benchmark. The method is zero-shot, model-agnostic and lightweight, offering a practical recipe for high-fidelity augmentationunder data scarcity. Our implementation is available at https://github.com/cactusYuri/llmbased_ASAP_AUG_main .