KWM-B: Key-Information Weighting Methods at Multiple Scale for Automated Essay Scoring with BERT
Tengteng Miao, Dong Roman Xu · Electronics · 2025
The Automatic Essay Scoring (AES) task aims to automatically evaluate the quality of papers through machines, thereby reducing the burden on teachers and improving the fairness of grading. Although pretrained models perform well in natural language processing (NLP) tasks, in the AES field, large pretrained language models like BERT have not shown greater advantages than other deep learning models such as CNN. Existing research has often analyzed articles as a whole, ignoring the differences in the importance of information in each part of the article. In response to this issue, this article proposes a new framework based on BERT, which adopts a multi-scale key information weighting method, focusing on important information at the token, sentence, paragraph, and document scales to express the semantic content of core ideas more effectively. In addition, this article enhances the model’s ability to identify key information through data augmentation techniques. The evaluation results using quadratic weighted kappa (QWK) indicate that the framework outperforms existing mainstream models on the Public Automatic Student Assessment Award (ASAP) dataset.