Research On Poetry Generation Problem Based On Improved Sequence Generation Adversarial Network

Chinese ancient poetry,with its refined language,rich imagination and sincere emotion,has been sung for thousands of years.It is not only the poet's records of life and expression of his emotions,but also his thinking of everything in the world and the fate of life.It has profound philosophical content.In recent years,with the rapid development of Internet technology,artificial intelligence has once again stepped onto the historical stage with a new attitude.Among them,poetry generation technology in natural language processing is a very challenging work.At present,although neural network technology has been used to study the generation of poetry at home and abroad,the generated poetry has the problem of theme deviation and unclear meaning because of the free model.In addition,how to make the machine-generated poetry more close to human creation and how to meet the requirements of poetry rhythm are research difficulties.In this thesis,the following improvements are made on the basis of sequence generative adversarial nets:(1)The encoding and decoding model based on attention mechanism is used to replace the long short-term memory model used in the generator part of the original model.In view of the fact that the input data of this thesis are several independent keywords,rather than a complete sequence with temporal information.Using a single long short-term memory model often results in poetry texts that are not strongly related to the keyword information.Therefore,this thesis proposes a encoding and decoding model based on attention mechanism.Firstly,the key words are extracted based on the semantic feature coding of the decomposing machine,and then a long short-term memory network is used as the decoder part of the model.There is also an attention module between the encoder and decoder,which keeps the key information and eliminates the unimportant information.(2)Learned from the conditional generative adversarial nets,the rhythm information of poetry is added to the input layer of the discriminator model,so that when the discriminator judges whether the text comes from the real sample or the generated sample,it also pays attention to whether the text meets the rhythm requirements of poetry.(3)In view of the problem that the original model uses Monte Carlo tree search method to complete the text,causing slow convergence of the model,this thesis proposes to use beam search instead of the original search method,and adds a penalty factor on the basis of the original scoring function,which plays a role in punishing the words that reduce the quality of the text.Experimental results show that the score of the improved sequence generation network is higher than that of the original model,from 0.739 to 0.803,which proves the effectiveness of the improved model.