A Survey of Deep Learning Applied to Story Generation

Abstract

With the rapid development of deep learning in the fields of text abstraction and dialogue generation, researchers are now reconsidering the long-standing story-generation task from the 1970s. Deep learning methods are gradually being adopted to solve problems in traditional story generation, making story generation a new research hotspot in the field of text generation. However, in the field of story generation, the widely used seq2seq model is unable to provide adequate long-distance text modeling. As a result, this model struggles to solve the story-generation task, since the relation between long text should be considered, and coherency and vividness are critical. Thus, recent years have seen numerous proposals for better modeling methods. In this paper, we present the results of a comprehensive study of story generation. We first introduce the relevant concepts of story generation, its background, and the current state of research. We then summarize and analyze the standard methods of story generation. Based on various divisions of user constraints, the story-generation methods are divided into three categories: the theme-oriented model, the storyline-oriented model, and the human-machine-interaction–oriented model. On this basis, we discuss the basic ideas and main concerns of various methods and compare the strengths and weaknesses of each method. Finally, we finish by analyzing and forecasting future developments that could push story-generation research toward a new frontier.