Abstract
With the recent advances of neural models and natural language processing,automatic generation of classical Chinese poetry has drawn significantattention due to its artistic and cultural value. Previous works mainly focuson generating poetry given keywords or other text information, while visualinspirations for poetry have been rarely explored. Generating poetry fromimages is much more challenging than generating poetry from text, since imagescontain very rich visual information which cannot be described completely usingseveral keywords, and a good poem should convey the image accurately. In thispaper, we propose a memory based neural model which exploits images to generatepoems. Specifically, an Encoder-Decoder model with a topic memory network isproposed to generate classical Chinese poetry from images. To the best of ourknowledge, this is the first work attempting to generate classical Chinesepoetry from images with neural networks. A comprehensive experimentalinvestigation with both human evaluation and quantitative analysis demonstratesthat the proposed model can generate poems which convey images accurately.