Sentence Generation Method by Extension of MolGAN Using Sentence Graph

Abstract

Deep learning solves many classification problems. However, it is difficult to solve problems with imbalanced data. Therefore, the data volume is increased for the purpose of balancing. This is called data augmentation. Generally, the method of image data augmentation uses noise addition, rotation, and the like. Recently, images are generated using the generative adversary network: GAN. However, data augmentation methods are difficult in natural language processing. In addition, manual data augmentation is burdensome and requires mechanical methods. Mechanical text augmentation is more difficult than images. Because it is difficult to analyze the feature of sentences. This paper proposes a sentence generation method by machine learning focusing on graph information. The graph information obtained by CaboCha is processed by graph Convolution. The proposed GAN was used to generate sentences, and then three experiments were performed to evaluate its effectiveness.