Poetry generation is a typical and popular task in natural language generation. While prior works have shown success in controlling either semantic or metrical aspects of poetry generation, there are still challenges in addressing both perspectives simultaneously. In this paper, we employ the Diffusion model to generate poetry in Sonnet and SongCi in Chinese for the first time to tackle such challenges. Different from autoregressive generation, our PoetryDiffusion model, based on Diffusion model, generates the complete sentence or poetry by taking into account the whole sentence information, resulting in improved semantic expression. Additionally, we incorporate a novel metrical controller to manipulate and evaluate metrics (format and rhythm). The denoising process in PoetryDiffusion allows for gradual enhancement of semantics and flexible integration of the metrical controller. Experimental results on two datasets demonstrate that our model outperforms existing models in terms of semantic, metrical and overall performance.

翻译：诗歌生成是自然语言生成中一项典型且热门的研究任务。尽管已有工作在控制诗歌生成的语义或格律方面取得了成功，但同时在两个维度上进行调控仍面临挑战。本文首次采用扩散模型生成中文十四行诗和宋词以应对上述挑战。与自回归生成不同，我们的PoetryDiffusion模型基于扩散机制，通过整合整句信息来生成完整句子或诗歌，从而提升语义表达能力。此外，我们引入了一种新颖的格律控制器，用于操控与评估格式和韵律等指标。PoetryDiffusion中的去噪过程使得语义得以逐步增强，并支持格律控制器的灵活集成。在两个数据集上的实验结果表明，我们的模型在语义、格律及综合性能上均优于现有模型。