In recent years much effort has been devoted to applying neural models to the task of natural language generation. The challenge is to generate natural human-like text, and to control the generation process. This paper presents a task-agnostic survey of recent advances in neural text generation. These advances have been achieved by numerous developments, which we group under the following four headings: data construction, neural frameworks, training and inference strategies, and evaluation metrics. Finally we discuss the future directions for the development of neural text generation including neural pipelines and exploiting back-ground knowledge.

翻译：近年来，大量研究致力于将神经模型应用于自然语言生成任务。其挑战在于生成类似人类自然语言的文本，并对生成过程进行控制。本文对神经文本生成的最新进展进行了任务无关的综述。这些进展源于众多研究突破，我们将其归纳为以下四大类：数据构建、神经框架、训练与推理策略，以及评估指标。最后，我们探讨了神经文本生成的未来发展方向，包括神经流水线技术及背景知识的利用。