Research And Application Of Deep Neural Networks In Generating Classical Chinese Poetry From Images

Classical Chinese poetry,as the essence of Chinese literature,has always been a very important part of Chinese culture.During the thousands of years in China,mil-lions of classical Chinese poems have been written to praise beautiful scenery,love and patriotism,etc.Classical Chinese poetry is still fascinating us today with its concise structure,rhythmic beauty and rich emotions.With the recent advances in neural models and natural language processing,au-tomatic generation of classical Chinese poetry with machine learning has drawn sig-nificant attention.Previous works mainly focus on generating poetry given keywords.However,because of the limited representation ability of text,several keywords often cannot convey users’ writing intent completely.Compared with keywords,image has more powerful representation ability because of rich semantic information and visual information in it and is more suitable for users’ writing intent.Therefore,generating classical Chinese poetry is a research topic worth studying.Generating classical Chinese poetry from images is much more challenging than generating classical Chinese poetry from text,since image contains very rich visual information which cannot be described completely using several keywords,and a good poem should convey the image accurately.In this dissertation,we propose a memory based neural network to automatically generate classical Chinese poetry from images,which exploits visual information and semantic information effectively.To the best of our knowledge,this is the first work attempting to generate classical Chinese poetry from images with neural networks.A comprehensive experimental investigation with both human evaluation and quantitative analysis demonstrates that the proposed model can generate poems which convey images accurately.The main contributions of this dissertation are:(1)We explore a new research topic of generating classical Chinese poetry from images.(2)We employ memory networks to tackle the problems of topic drift while resolving the restriction of requiring a limited number of keywords.(3)We integrate semantic information and visual information in images when generating poems.