AI for Poetry

Description

Meter, rhyme, and stylistic constraints for AI-generated poetry, with evaluation and human-AI curation.

When to use

You want to generate or co-write poems under formal constraints such as meter, rhyme, syllable counts, or a specific literary style.

Key concepts

• Formal constraints: meter, rhyme scheme, syllable counts, stanza forms, and refrain patterns.
• Poetic style prompting: persona, era, mood, imagery, alliteration, and lexical register.
• Controllable generation: constrained decoding, iterative refinement, and rule-based post-processing for rhyme and meter.
• Evaluation: automatic metrics, LLM-as-judge, and human evaluation for novelty, emotion, and aesthetic quality.
• Ethics and attribution: respect public-domain or licensed training data and credit human curators.

Code pattern

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "openai-community/gpt2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

prompt = """Write a Shakespearean sonnet about autumn.
Use iambic pentameter and an ABAB CDCD EFEF GG rhyme scheme:

"""

inputs = tokenizer(prompt, return_tensors="pt")
output = model.generate(
    **inputs,
    max_new_tokens=200,
    temperature=0.9,
    do_sample=True,
)
poem = tokenizer.decode(output[0], skip_special_tokens=True)
print(poem)

Tuning notes

• Use constrained decoding or post-hoc rhyme checking to satisfy form.
• Fine-tune on poetry corpora for stronger stylistic control.
• Balance novelty with readability; avoid cliches.
• Validate meter and rhyme with dedicated tools such as pronouncingpy or pyphen.

Verification

1. Generate 10 sonnets and check rhyme and meter compliance automatically.
2. Compare human vs. AI poems in a small blinded preference test.
3. Evaluate diversity and novelty across a themed set of poems.

References

• https://doi.org/10.1613/jair.1.20584
• https://aclanthology.org/W17-3502/
• https://aclanthology.org/2024.emnlp-main.1097/
• https://computationalcreativity.net/iccc24/papers/ICCC24_paper_164.pdf