基于多对抗训练的古诗生成方法  

Chinese poetry generation model with multi-adversarial training

作　　者：黄文明 任冲[1] 邓珍荣 Huang Wenming;Ren Chong;Deng Zhenrong(College of Computer&Information Security Guilin University of Electronic Technology,Guilin Guangxi 541000,China;Guangxi Colleges&Universities Keys Laboratory of Cloud Computing&Complex Systems,Guilin University of Electronic Technology,Guilin Guangxi 541000,China)

机构地区：[1]桂林电子科技大学计算机与信息安全学院,广西桂林541000 [2]桂林电子科技大学广西高校云计算与复杂系统重点实验室,广西桂林541000

出　　处：《计算机应用研究》2021年第1期164-168,共5页Application Research of Computers

基　　金：广西自然科学基金资助项目(2018GXNSFAA138132);桂林电子科技大学研究生教育创新计划资助项目(2019YCXS050)。

摘　　要：目前许多古诗生成方法离人类创作的水平仍有较大的差距,尤其是在主题关联性及诗句的语义方面。为弥补现有方法的不足,提出一种多对抗训练的古诗生成框架。以融合了注意力机制并采用双编码器的序列到序列模型作为古诗生成器,以层级RNN和TextCNN组合的多判别模型指导古诗的生成,同时基于策略梯度进行多对抗训练。在古诗意象数据集上进行实验表明,相较于已提出的方法,基于多对抗训练的古诗生成方法有效提升了诗句与意象词之间的关联性,古诗所表现的语义内涵也更加丰富。Many generation methods still have a large gap from the human creation,especially on the aspects of topical relevance and semantics of verses.To address these shortcomings of existing methods,this paper proposed a framework for generating poems with multiple adversarial training.The poetry generator used the sequence-to-sequence model with the attention mechanism and dual-encoding.Two discriminative model guided the poetry generation,included hierarchical RNN and TextCNN.Meanwhile,the framework used policy gradient for multi-adversarial training.Experiments show that the poetry gene-ration method based on multi-adversarial training effectively improves the relevance between verses and vision keywords,and the semantic connotation of poem is more abundant.

关 键 词：序列到序列 注意力机制 对抗学习 古诗生成 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]