Effectiveness of Data Augmentation in Pointer-Generator Model

抄録

We propose a new data augmentation method in automatic summarization system, especially the Pointer-Generator model. A large corpus is required to create an automatic summarization system using deep learning. However, in the field of natural language processing, especially in the field of automatic summarization, there are not many data sets that are sufficient to train automatic summarization system. Therefore, we propose a new method of data augmentation. We use the Pointer-Generator model. First, we determine the importance of each sentence in an article using topic model. In order to augment the data, we remove the least important sentence from an input article and use it as a new article. We examine the effectiveness of our proposed data augmentation method in automatic summarization system.