An Algorithmic Study On Chinese Classical Poetry Composition With Literary Expressiveness

Natural Language Generation(NLG)is an essential branch of Natural Language Pro-cessing(NLP).As a highly literary and artistic form of language,poetry has a far-reaching influence on the development of human culture and society.Among various poetry gen-res,Chinese classical poetry,with its concise expressions and regular forms,as well as colourful content and delicate feelings,could be an ideal entry point of investigating NLG.Research on the automatic composition of Chinese classical poetry could date back to the end of the 19 th century and has gradually become another research focus due to its con-siderable research value.From the perspective of explorations,this task could promote the study of human writing mechanism and the construction of computational creativity?From the perspective of applications,this task could benefit a wide range of products,such as entertainments,intelligent assistants of education,and literary research.Traditional methods usually regard automatic poetry composition as a kind of se-quence prediction/mapping task but neglect poetry's literary properties.Such a practice,for one thing,leads to context inconsistency and poor topic relevance of generated poetry?for another,it makes generated poems indistinguishable from each other,which hurts nov-elty and interest.Such defects cause a poor user experience of these systems and severely limit their effectiveness in downstream applications.To tackle these problems,we make an effort to improve the literary expressiveness of generated poems.To this end,we handle two critical components of literary expressiveness,namely textual quality and aesthetic feature,and systematically propose solutions for each corresponding research challenge.For textual quality,we mainly consider context coherence and topic relevance of generated poems.For better context coherence,we innovatively propose two methods,Salient Clue Mechanism and Working Memory Model,to eliminate noise in context dur-ing the generation process of a poem,enhancing the relevancy of different lines and the consistency of theme and topic.To improve topic relevance,we design a separate topic memory and adopt text style transfer techniques to handle multiple keywords and a whole sentence inputs respectively,making the generated poetry closely related to user inputs.For aesthetic features,we focus on the novelty and stylization of generated poems.To promote novelty,we model and quantify each human evaluation criterion to encourage the model to write humanly.To achieve stylization,we devise a novel Mixed Latent Space method to endow generated poems with distinctive and controllable styles.Furthermore,we have implemented all methods mentioned above and integrated them into our online Chinese classical poetry composition system,Jiuge.Since its re-lease,our system has been used more than ten million times,having a profound and pos-itive impact on both the academic community and the public.