Dynamic Inspiring Sets for Sustained Novelty in Poetry Generation

One activity recognised as an interesting instance of creativity is the ability of poets to systematically come up with new poems, irrespective of how many they have already written in the past. Poets who periodically produce a new poem, different from the earlier ones and comparatively valuable, are considered to embody a more interesting type of creativity than poets who have produced only one good poem in their lifetime, or those that produce a succession of poems all built following a standard recipe so obvious that it can almost be described as a template. This paper explores the constraints imposed on computational creative processes by the requirement of novelty with respect to previous outputs. Two issues emerge as fundamental: how to evaluate novelty against a set of artifacts taken as reference, and how to adjust construction procedures so that each successive run leads to significantly different output. A possible modelling of these two issues is proposed in terms of two sets of sample artifacts: a reference set (against which novelty is measured) and a learning set (used to configure the construction procedures). Over this basic model, extended discussion is carried out to draw out interesting insights for the design of computational creative systems.