An evolutionary algorithm approach to poetry generation
Poetry is a unique artifact of the human language faculty, with its defining feature being a strong unity between content and form. Contrary to the opinion that the automatic generation of poetry is a relatively easy task, we argue that it is in fact an extremely difficult task that requires intelligence, world and linguistic knowledge, and creativity. We propose a model of poetry generation as a state space search problem, where a goal state is a text that satisfies the three properties of meaningfulness, grammaticality, and poeticness. We argue that almost all existing work on poetry generation only properly addresses a subset of these properties. In designing a computational approach for solving this problem, we draw upon the wealth of work in natural language generation (NLG). Although the emphasis of NLG research is on the generation of informative texts, recent work has highlighted the need for more flexible models which can be cast as one end of a spectrum of search sophistication, where the opposing end is the deterministically goal-directed planning of traditional NLG. We propose satisfying the properties of poetry through the application to NLG of evolutionary algorithms (EAs), a wellstudied heuristic search method. MCGONAGALL is our implemented instance of this approach. We use a linguistic representation based on Lexicalized Tree Adjoining Grammar (LTAG) that we argue is appropriate for EA-based NLG. Several genetic operators are implemented, ranging from baseline operators based on LTAG syntactic operations to heuristic semantic goal-directed operators. Two evaluation functions are implemented: one that measures the isomorphism between a solution’s stress pattern and a target metre using the edit distance algorithm, and one that measures the isomorphism between a solution’s propositional semantics and a target semantics using structural similarity metrics. We conducted an empirical study using MCGONAGALL to test the validity of employing EAs in solving the search problem, and to test whether our evaluation functions adequately capture the notions of semantic and metrical faithfulness. We conclude that our use of EAs offers an innovative approach to flexible NLG, as demonstrated by its successful application to the poetry generation task.