Zero-shot Sonnet Generation with Discourse-level Planning and Aesthetics Features
This work addresses the challenge of creative language generation for poetry enthusiasts and researchers, offering a zero-shot approach that avoids the need for large poem datasets, though it is incremental in its application of existing techniques to a specific domain.
The authors tackled the problem of generating sonnets without training on poetry data by designing a hierarchical framework that includes discourse-level planning, rhyme generation, and aesthetic polishing, resulting in sonnets that were rated as more coherent, poetic, and creative than strong baselines in evaluations.
Poetry generation, and creative language generation in general, usually suffers from the lack of large training data. In this paper, we present a novel framework to generate sonnets that does not require training on poems. We design a hierarchical framework which plans the poem sketch before decoding. Specifically, a content planning module is trained on non-poetic texts to obtain discourse-level coherence; then a rhyme module generates rhyme words and a polishing module introduces imagery and similes for aesthetics purposes. Finally, we design a constrained decoding algorithm to impose the meter-and-rhyme constraint of the generated sonnets. Automatic and human evaluation show that our multi-stage approach without training on poem corpora generates more coherent, poetic, and creative sonnets than several strong baselines.