A chart generation system for topical metrical poetry

Several poetry generation systems that are in some way inspired or motivated by existing articles such as newspaper stories have recently appeared. However, most if not all of them employ template-based generation, which limits both the expressiveness of the system and the ability to faithfully convey the message of the source article. In this paper we present our work on a poetry generation system that uses a dependency parser to extract the predicate argument structure of the input article, and tries to maintain this structure through deep syntactic text generation whilst complying with a given target form. The combinatorial nature of this task presents huge challenges, and we describe several improvements that have been applied in an attempt to produce poetry in a tractable fashion.