A Study Of Chinese Quatrain Generation Based On Deep Learning Methods

Poetry is a condensed and special literary form.Traditional Chinese poetry,as an important cultural heritage of our country,embodies the wisdom of our working people.The quatrain is a representative poetry genre in traditional Chinese poetry,which has strict requirements in terms of structure,flatness and rhyme.Creating a qualified quatrain is not an easy task for ordinary people,but for computers,how to automatically generate quatrains is also a challenging research topic.Research on the automatic generation of quatrains can lower the threshold of poetry creation,let ordinary people feel the charm of poetry creation,and help the inheritance of traditional Chinese culture;The impact of traditional poets and poetry researchers will,to a certain extent,promote the innovation and development of traditional Chinese poetry;meanwhile,quasi-sentence generation is a special and interesting research in the field of natural language processing,which can inspire other text types to generate research and promote Development of language processing related technologies.Therefore,the research on the generation of quatrains is scientific and feasible.The research on the generation of quatrains has gone through three stages: the generation method based on rules and templates,the method based on statistical machine learning and the method based on deep learning.The poems generated by the first two methods usually require manual participation,and often have lower-level errors.With the continuous development of deep learning technology,deep learning-based methods have performed well in poem generation and become the mainstream.In this paper,based on the existing generation methods of quatrains,we propose a Keyword Transformation and Expansion Quatrain Generation Model(KTEQG)to solve the problems of topic drift and semantic incoherence in the generation of quatrains.The model divides the generation of quatrains into three stages: keyword conversion,keyword expansion and quatrains generation.First,extract the unique keywords of the user's writing intent text and perform keyword classical Chinese word conversion,then expand based on the converted topic keywords,assign topic keywords to each sentence,and finally based on the encoder-decoding of the attention mechanism The model uses the assigned topic keywords and historically generated content as input to generate quasi-sentences.In addition,this paper aims at the shortcomings of the current poetry evaluation methods in terms of topic,mood and emotion,and combines the literary characteristics ofquatrains to improve the existing manual evaluation methods and establish a more scientific manual evaluation system.This study is a comparative experiment with the mainstream quatrains generation model.After BLEU automatic evaluation and manual evaluation,the experiment shows that the utterances generated by the KTEQG model proposed in this paper have better performance in metric,topic and content expression.The model was also subjected to a Turing test.The experiment showed that the KTEQG model has the intelligence of normal human beings,and has reached the level of ordinary humans in the creation of quatrains.