Creative Language Generation in a Society of Engagement and Reflection

Abstract

Conference proceeding from ICCC'20 International Conference on Computational Creativity.

Many existing models of narrative and language generation use rigid sequences of steps which are cognitively implausible and limit creativity. Iterative models based on Sharples' cycle of engagement and reflection improve on this by incorporating self-evaluation but still have a rigid arrangement of parts. This paper outlines how a multi-agent approach could be used to break apart the cycle into a more fluid society of engagement and reflection, whose constituent agents interact with one another to produce a text. Our approach is to work in a simple domain in order to focus on the underlying processes, and to avoid the Eliza effect during evaluation.