Short Text Generation Based on Adversarial Graph Attention Networks

Abstract

Text generation has attracted more and more attention in the field of natural language. Recently, GAN (Generative Adversarial Networks) have been widely used in text generation, among which the GAN-based models, such as SeqGAN and SentiGAN, have shown remarkable effects in text generation. However, previous text generation models simply use CNN (Convolutional Neural Networks) as discriminators and ignore relationships between the same-label texts. Meanwhile, most models only consider using a single generator to generate a single species text, not for multispecies texts. To meet the requirements, in this paper, we propose a novel framework model-SGATGAN, which applies GAT (Generative Attention Nets) as the discriminator to establish the connection between the texts of the same type. It also provides a method of generating multispecies texts using a single generator. In this model, the graph attention neural network is used as the discriminator via the feedback to guide the generator in a specific location to generate a specific type of short text. Experimental results on two benchmarks show that our model significantly outperforms previous methods, giving state-of-the-art results in short text generation.