부산대학교 도서관

To address the low-resource problem of machine learning tasks, including natural language processing, the effectiveness of data augmentation is well known. In recent natural language processing, data augmentation based on paraphrase generation has been used successfully in many applications. However, unlike English and Chinese, there is no large-scale corpus for training paraphrase generation models in Japanese, and thus data augmentation based on paraphrase generation is not available for Japanese. We release JParaBank, a large-scale Japanese paraphrase corpus of 21 million sentence pairs. Experimental results on the JGLUE benchmark show that data augmentation by paraphrase generation using JParaBank improves performance on many tasks.