Research On Automatic Generation Of Poetry And Couplet Based On Neural Network

Poetry and couplets are unique artistic forms of Chinese language.They are rich in semantics and orderly in rhythm.They have a strong sense of beauty and are loved by people.In NLP,the automatic generation of poetry and couplet is challenging.Realizing the automatic generation of poetry and couplet has great significance for the promotion of Chinese traditional culture.This paper studies the automatic generation of poetry and couplet based on neural network and multi-task learning.The main work is as follows:First,aiming at improving the problem that computer-written poetry has an unclear theme,the content of poetry and writing intention are inconsistent,this paper learns from the planning-based poetry generation and imitates the process of writing poetry by poet,proposed a poetry generation method based on seq2 seq model.The method divides poetry generation into two stages.In the first stage,a keyword extension model based on attention mechanism is used to obtain the outline of poetry according to user's input.In the second stage,a seq2 seq model with double encoders and attention mechanism is used to generate poetry according to the outline.Experiments show that compared with the benchmark method,the poetry is generated by the proposed method has a clearer theme,and the content of poetry is more consistent with writing intention.Second,in Chinese traditional literature,there are many similarities between poetry and couplet,and their automatic generation methods are roughly same for computer.Aiming at this phenomenon,this paper proposes a multi-task learning method for automatic generation of poetry and couplet,which is based on the previous poetry generation method.The method also constructs the outline according to the user's input,and then uses a novel multi-task learning model to generate poetry and couplet.In the multi-task learning model,the encoder parameters are shared and the decoder parameters are not shared.The encoder of model can learn both poetry and couplet features.The decoder of model retains the different features between poetry and couplet.Experiments show that compared with the benchmark method,the multi-task learning model has stronger generalization ability and better performance than the single-task learning model.The main innovations of this paper are as follows:(1)Based on the research of the planning-based poetry generation,designed a keyword expansion model based on attention mechanism for keyword expansion.(2)A multi-task learning model is designed for poetry and couplet generation,which has stronger generalization ability.(3)For the first time,writing intention is integrated into the generation model of couplet to generate personalized couplet.