Research On Poetry Generation Algorithm Based On Generative Adversarial Nets

Poetry has both language and artistic beauty,and it is an advanced expression of human wisdom and creativity.Chinese classical poetry is the best representative of our country's outstanding traditional culture and “Cultural confidence”.Traditional natural language processing techniques often use rules and templates,statistical mechanized translation and other methods to generate poetry,and the quality of works is poor.The application of deep learning technology,which has been rapidly developed and widely used in recent years,to the automatic generation of Chinese classical poetry is a very forward-looking and practical research.There are also quatrains and rhymes,five words and seven words in Chinese classical poetry.This article mainly aims at the generation of five words and quatrains.The main work of this article is summarized as follows:(1)Convert the poetry generation task into two related adversarial training subtasks,implement the poetry generator with Gated Recurrent Unit,and improve the generation adversarial nets single discriminator structure to a double discriminator structure,in which the topic discriminator will distinguishes the generate poetry to match,mismatch and generate three categories,the poetic discriminator distinguishes the generated poetry into four categories: poetic category,disordered category,paragraph category and generated category.The discriminant results of the two discriminators are combined as a reward to provide feedback to the poetry generator,and the quality of poetry generation is continuously improved to ensure that the generated poetry has a clear theme and beautiful poetry.(2)The two methods of automatic evaluation and manual evaluation are used to comprehensively evaluate the quality of the generated poems.The experiments show that the three automatic evaluation indexes of topic relevance,word novelty and BLEU are respectively improved by 0.233,0.126 and 0.018 compared with the RNNPG model,and respectively improved by 0.111,0.091 and 0.024 compared with the PPG model.The three manual evaluation indexes of relevance,logic and expansibility are respectively improved by0.68,0.5 and 0.52 compared with the RNNPG model.Turing test also show that the gap between the poems generated by the model in this paper and human poems is significantly reduced.(3)The copyright of poems generated by artificial intelligence depends on the originality of the works.Experiments show that the originality of artificial intelligence poems mainly depends on the collection of poetry corpus and the creation of artificial intelligence algorithm models.The contribution to the originality of the work is different.This article believes that the copyright ownership of artificial intelligence poetry works can be further determined to be shared by poetry corpus collectors and designers of artificial intelligence algorithm models.