多样性文本生成方法研究

Research on Diverse Text Generation

多样性文本生成是自然语言处理领域的重要研究之一,旨在让机器生成符合语法规范的、多样的、相关的文本。由于广泛用于文本生成任务的序列到序列模型(Seq2Seq)并不适合生成多样性文本,现有研究以重新构造模型优化目标与改变Seq2Seq模型结构为切入点,分别提出最大化互信息、变分自编码(VAE)、条件变分自编码(CVAE)等方法,然而VAE/CVAE之类的方法仍旧面临着极大的优化挑战。

Diverse text generation is one of the essential tasks in natural language processing,which aims to generate varied and relevant texts that conform to grammatical norms.Since the Sequence-to-Sequence(Seq2Seq)model,which generally used in text generation tasks,is not suitable for generating diverse outputs,the existing researches focus on either reconstructing the optimization objectives or changing the structure of Seq2Seq,such as mutual information objectives and Variational Auto-encoder(VAE)/Conditional Variational Auto-encoder(CVAE).However,VAE/CVAE still face optimization challenges.