基于深度强化学习的文本生成研究综述

Review of Text Generation Based on Deep Reinforcement Learning

文本生成任务需要对大量词汇或语句进行表征,且可将其建模为序列决策问题.鉴于深度强化学习(deep reinforcement learning,DRL)在表征及决策方面的优良性能,DRL在文本生成任务中发挥了重要的作用.基于深度强化学习的文本生成方法改变了以最大似然估计为目标的训练机制,有效解决了传统方法中存在的暴露偏差问题.此外,深度强化学习和生成对抗网络的结合进一步提高了文本生成质量,并已取得了显著的成果.本综述将系统阐述深度强化学习在文本生成任务中的应用,介绍经典模型及算法,分析模型特点,探讨未来深度强化学习与文本生成任务融合的前景和挑战.

Text generation tasks require representation of a large number of words or statements and can be modeled as sequential decision problems.In view of the excellent performance of deep reinforcement learning in representation and decision-making,it plays an important role in text generation tasks.The text generation method based on deep reinforcement learning changes the training mechanism aiming at maximum likelihood estimation and effectively solves the problem of exposure bias in traditional methods.In addition,the combination of DRL and generative adversarial networks has improved the quality of text generation and has achieved remarkable results.This review will elaborate the application of DRL in text generation tasks,introduce the classical models and algorithms,analyze the characteristics of the models,and discuss the prospects and challenges of the future integration of DRL and text generation tasks.