A Multi-perspective Automatic Scoring Method Using Multiple LLMs in an e-learning System for Summary Essays

This study aims to develop an e-learning system for summary essay tasks. In summary problems, a wide variety of answers can be regarded as correct; therefore, the scoring function must be capable of multi-perspective evaluation. The proposed system seeks to realize an automatic scoring function that enables multi-perspective evaluation by utilizing generative AI. To achieve multi-perspective evaluation, we conducted experiments using multiple generative AI models and found that their evaluation policies differed depending on the model. In the learning process, it is assumed that learners will repeatedly receive scoring results and feedback comments; however, inconsistent evaluation policies may cause confusion among learners. To address this issue, we propose a method in which multiple generative AI models first generate evaluation criteria, and then the scoring is performed according to these criteria. By selecting the evaluation criteria that align most closely with their own perspectives, the system enables multi-perspective evaluation while maintaining consistency in feedback for learning improvement.