Cross-Prompt Automated Essay Scoring via Reinforcement Learning-Based Data Valuation
Takumi Shibata, Masaki Uto · IEEE Access · 2025
Automated essay scoring (AES) aims to automatically grade essays, thereby reducing the time and cost associated with manual scoring. The most common AES methods are classified under the prompt-specific approach, which involves developing a scoring model exclusively for a target prompt by using a dataset of scored essays corresponding to that prompt. Meanwhile, recent studies have emphasized the cross-prompt approach. This method leverages scored essay data from other prompts, referred to as source prompts, to build an AES model for the target prompt. However, these cross-prompt methods have limitations in that they do not consider the presence of source-prompt essays that can potentially have a negative impact on the construction of the AES model for the target prompt. To address this limitation, we propose a novel cross-prompt AES method that utilizes data valuation based on reinforcement learning (DVRL). The proposed method enables the value-driven use of source-prompt essays, which positively contributes to improving the scoring accuracy of the AES for the target prompt. Experiments on a benchmark dataset demonstrate that the proposed method enhances the performance of various AES models in cross-prompt scoring settings.