Research map / Image generation / Autoregressive and latent generation
PoetryDiffusion: Towards Joint Semantic and Metrical Manipulation in Poetry Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PoetryDiffusion, a continuous diffusion language model with a separately trained BERT-based metrical controller that steers denoising to jointly control semantics and format/rhyme/tone in sonnet and SongCi generation.
Problem
Controllable poetry generation requires both coherent semantics and strict metrical rules (line format, rhyme, tone), which prior autoregressive, GAN/VAE, and LLM approaches handle only partially.
Approach
Apply a continuous diffusion model over BERT word embeddings for whole-poem semantics, and attach a separately trained metrical classifier (format, rhyme, tone) whose loss is injected at every denoising step so metrics are enforced during generation without concatenating control codes to the content encoder.
Key results
- On Sonnet, PoetryDiffusion reaches overall score 67.78 vs 67.31 for Llama2-70B-chat and 64.30 for ChatGPT, with format 100.00 and rhyme 52.28.
- On SongCi, PoetryDiffusion achieves overall 76.43 with format 99.51, tone 91.64, rhyme 95.37, but MRCG has higher metrical average (97.02 vs 95.51).
- In human evaluation, PoetryDiffusion has the best average (3.43 Sonnet, 3.46 SongCi), with ChatGPT higher on fluency and meaningfulness.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the metrical controller (ablation) | PoetryDiffusion · Sonnet test | semantic average: 61.70 -> 63.60 | Helped |
| add the metrical controller (ablation) | PoetryDiffusion · Sonnet test | metrical average: 59.84 -> 76.14 | Helped |
| add the metrical controller (ablation) | PoetryDiffusion · Sonnet test | overall: 61.08 -> 67.78 | Helped |
| add the metrical controller (ablation) | PoetryDiffusion · SongCi test | semantic average: 56.35 -> 57.35 | Helped |
| add the metrical controller (ablation) | PoetryDiffusion · SongCi test | metrical average: 65.24 -> 95.51 | Helped |
| add the metrical controller (ablation) | PoetryDiffusion · SongCi test | overall: 60.79 -> 76.43 | Helped |
| use PoetryDiffusion instead of Llama2-70B-chat | Sonnet generation · Sonnet test | overall: 67.31 -> 67.78 | Helped |
| use PoetryDiffusion instead of MRCG | SongCi generation · SongCi test | overall: 75.72 -> 76.43 | Helped |
Surprising findings
- On SongCi, the MRCG baseline achieves slightly higher metrical scores than PoetryDiffusion, which the authors attribute to MRCG's forced word replacement at the end.
- Removing the metrical controller lowers Sonnet rhyme accuracy from 52.28 to 23.68 while semantic scores also drop slightly.
Limitations
- Metrical scores on SongCi are slightly below MRCG's.
- Human evaluation uses only 25 sonnets and 25 SongCi poems and a single baseline comparison set.
- ChatGPT outperforms the model on fluency and meaningfulness in human ratings.
Experimental setup
Datasets: Sonnet (3,355 sonnets) SongCi (82,724 poems)
Benchmarks: Sonnet SongCi
Metrics: BLEU ROUGE Distinct PPL format accuracy rhyme accuracy tone accuracy human ratings (fluency, coherence, meaningfulness, poeticnes
Baselines: SeqGAN MRCG SongNet GPT3 ChatGPT Llama2-70B-chat
Builds on: Diffusion-LM DDIM SongNet MRCG
Closest papers by meaning
- Energy-Based Diffusion Language Models for Text Generation (ICLR 2025 · similarity 0.86)
- Scaling Diffusion Language Models via Adaptation from Autoregressive Models (ICLR 2025 · similarity 0.85)
- S²MILE: Semantic-and-Structure-Aware Music-Driven Lyric Generation (AAAI 2025 · similarity 0.85)
- Meta-Diffu$B$: A Contextualized Sequence-to-Sequence Text Diffusion Model with Meta-Exploration (NeurIPS 2024 · similarity 0.85)
- SongBloom: Coherent Song Generation via Interleaved Autoregressive Sketching and Diffusion Refinement (NeurIPS 2025 · similarity 0.85)
- Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.