The idea to personalize technology is not novel but despite impressive progress of technologies in recent years, human-machine interfaces are still fairly deprived of any distinct personality features. We believe that one way to make such interfaces more human-friendly and avoid the uncunny valley is to stylize the generated texts so that they resemble the actual style of a communication of an actual person. Poetry is a well structured type of text data that also is known to have distinguished stylistic features characterizing different authors. That is why we have decided to specifically work with poetry.
We have formulated a problem of stylized text generation and propose a method capable of solving this problem in a multi language setup. A version of a language model based on a long short-term memory artificial neural network with extended phonetic and semantic embeddings is used for stylized poetry generation. The quality of resulting poems generated by the network are estimated both objectively through cross entropy with samples of original texts and subjectively through a survey. Humans attribute machine generated texts to the target author almost as frequently as they attribute original texts to the author in question.