Japanese-English Model Integration and GPT-4-Based Data Augmentation for Japanese Debate Quality Assessment

Recently, the incorporation of group debates has emerged as a strategic approach for measuring communication skills within the realms of entrance and employment examinations. However, the equitable evaluation of debates for such settings proves challenging. Therefore, a system is required to assess debates appropriately. In this study, we work on the quality assessment of debates focused on the Japanese language. The studies related to debate primarily focused on English, while those targeted at Japanese are limited. Consequently, the availability of debate data in Japanese is more constrained compared to English. To handle low-resource data, we propose two methods: Japanese-English model integration and data augmentation using GPT-4. We show that Japanese-English model integration does not yield statistically significant scores, while data augmentation using GPT-4 enhances assessment performance.