Poetry Translation With Meter and Rhyme

ISEF Category: Technology Enhances the Arts

Subcategory: Human Information Exchange  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

Most machine translators flatten poetry into plain prose. That means the rhythm, rhyme, and music can vanish in a single pass. Your project asks a harder question, can a translator keep the meaning and the poem’s structure at the same time?

What Is It?

This project studies a translation system that tries to keep poetry sounding like poetry. Normal translation tools focus on meaning. Your version also tracks meter, which is the beat pattern in a poem, and rhyme, which is the sound at the end of lines.

Think of it like translating a song while keeping the melody. The system explores many possible word choices, then ranks them by meaning, rhythm, syllable count, and rhyme. A beam search is a method that keeps the best few candidate translations instead of only one, so the model can compare options before it commits.

You can test this on sonnets in English, Spanish, and Hindi. Each language handles stress, syllables, and rhyme in different ways, so your project becomes a real test of whether poetry can survive translation across very different sound systems.

Why This Is a Good Topic

This is a strong science fair topic because you can measure more than one outcome. You can compare meaning, meter, rhyme, and human preference, then see which design choices help or hurt each one. It connects to real problems in machine translation, digital humanities, and multilingual access to literature. You also get to work with language data, model evaluation, and human ratings, which makes the project feel real and research-driven.

Research Questions

• How does adding meter constraints affect human ratings of translated sonnets?
• How does adding rhyme constraints affect the accuracy of translated meaning?
• Does a beam search with phonetic embeddings produce better rhythmic output than plain machine translation?
• To what extent do bilingual judges prefer translations that preserve iambic pentameter over translations that preserve literal meaning?
• Which language pair, English to Spanish, English to Hindi, or Spanish to English, produces the best balance of rhyme and meaning?
• What is the effect of syllable-count penalties on the consistency of line length across translated poems?

Basic Materials

• Laptop with access to a large language model or translation model interface.
• Public-domain sonnet texts in English, Spanish, and Hindi.
• Pronouncing dictionary or grapheme-to-phoneme resource for English.
• Hindi and Spanish syllable or pronunciation references.
• Spreadsheet software for rating and score tracking.
• Digital timer or note-taking app for judge sessions.
• Survey form tool for bilingual judge Likert ratings.

Advanced Materials

• GPU-equipped workstation or cloud compute access for model experiments.
• Open-source translation model checkpoints.
• Phonetic embedding library or phoneme encoder.
• Automatic syllable counter for multiple languages.
• Sentence alignment tool for parallel texts.
• Statistical analysis software for inter-rater agreement and significance testing.
• Annotation platform for bilingual judge review.

Software & Tools

• Python: Runs preprocessing, scoring, and comparison scripts for translation candidates.
• Hugging Face Transformers: Provides access to open-source translation and language models.
• NumPy: Handles numerical scoring for meter, rhyme, and judge data.
• pandas: Organizes poem pairs, ratings, and experiment results.
• Praat: Helps inspect speech sounds and phonetic patterns when you analyze rhyme and stress.

Experiment Steps

1. Define the exact translation goal, such as preserving meter, rhyme, meaning, or some weighted mix of the three.
2. Choose a small, controlled poem set so you can compare outputs line by line.
3. Build a scoring plan that converts syllable count, rhyme match, and semantic similarity into measurable variables.
4. Design the search strategy that ranks candidate translations before you generate final outputs.
5. Plan your judge study so bilingual raters score the same outputs with the same rubric.
6. Decide how you will compare human ratings against automatic metrics and simple baseline translators.

Common Pitfalls

• Treating literal meaning as the only success measure, which hides whether the poem still sounds poetic.
• Comparing translations from different source poems, which makes meter and rhyme scores hard to trust.
• Using rhyme matches without checking stress patterns, which can make the line sound forced.
• Mixing judge criteria in one rating item, which prevents you from telling rhythm problems from meaning problems.
• Ignoring language-specific pronunciation rules, which can break syllable counts in Spanish or Hindi.

What Makes This Competitive

A stronger project does more than report that one version sounds better. You would compare multiple search strategies, multiple scoring weights, and at least one baseline translator. You would also test agreement between automatic metrics and bilingual human judges. If your analysis finds a pattern, such as rhyme helping one language pair but hurting another, the project starts to look like real research instead of a demo.

Project Variations

• Translate haikus or limericks instead of sonnets and compare whether shorter forms are easier to preserve.
• Focus on one language pair, such as English to Spanish, and test several rhyme-preserving decoding rules.
• Replace bilingual judge ratings with expert literary analysis and compare human and automatic scores.

Learn More

• MIT OpenCourseWare, Introduction to Natural Language Processing: Search MIT OpenCourseWare for NLP lectures and assignments that explain translation models and evaluation.
• Stanford Online, Natural Language Processing with Deep Learning: Search the course materials for sequence models, attention, and decoding methods.
• PubMed: Search for review articles on computational linguistics, machine translation evaluation, and human judgment studies.
• ACL Anthology: Search for peer-reviewed papers on poetry translation, constrained decoding, and rhyme-preserving generation.
• NLTK Book: A free online textbook for language processing basics, including tokenization, phonetics, and text analysis.
• Princeton University, The Princeton Encyclopedia of Poetry and Poetics: Search your library or preview access for background on meter, rhyme, and poetic form.