Component-based approach to automatic poetry generation

This paper describes two approaches to generating poetic texts on a given topic, one of which we used while participating in ClassicAI, a contest in developing poetry generators in a specific style held by Sberbank. In the first one, we used topic modeling for extracting keywords that a certain topic is characteristic of, applied a text data augmenter to replace parts of the source of the poetry style with thematic words, and then applied a poetic consistency checker to maintain rhyme and rhythm in the output text. In the second one, we used semantic search for obtaining odd lines for the output texts and then phonetic search that selected lines similar in rhyme and rhythm to the given lines and used them as even ones. In this paper, we describe both of our approaches, analyze their benefits and weak spots, provide information on the results of the competition and suggest possible improvements