[Paper Review] Deep-speare: A Joint Neural Model of Poetic Language, Meter and Rhyme

Jey Han Lau, Trevor Cohn+3 more|arXiv (Cornell University)|Jul 10, 2018

Music and Audio Processing31 references4 citations

TL;DR

This paper proposes Deep-speare, a joint neural model that jointly generates poetic language, meter, and rhyme in Shakespearean sonnets using a unified architecture. Despite excelling in rhyme and meter—producing poems nearly indistinguishable from human writing—expert evaluation reveals that machine-generated poems still underperform in readability and emotional impact, suggesting future work should prioritize poetic substance over formal constraints.

ABSTRACT

In this paper, we propose a joint architecture that captures language, rhyme and meter for sonnet modelling. We assess the quality of generated poems using crowd and expert judgements. The stress and rhyme models perform very well, as generated poems are largely indistinguishable from human-written poems. Expert evaluation, however, reveals that a vanilla language model captures meter implicitly, and that machine-generated poems still underperform in terms of readability and emotion. Our research shows the importance expert evaluation for poetry generation, and that future research should look beyond rhyme/meter and focus on poetic language.

Motivation & Objective

• To develop a joint neural model that simultaneously captures poetic language, meter, and rhyme for sonnet generation.
• To assess the quality of generated poems using both crowdworker and expert evaluations.
• To investigate whether explicit modeling of meter and rhyme improves poetic quality beyond what a vanilla language model can achieve.
• To evaluate whether form-focused models compromise readability and emotional depth in generated poetry.
• To release a novel sonnet dataset and full source code for reproducibility and future research.

Proposed method

• Proposes a joint architecture combining a language model with separate modules for stress (meter) and rhyme prediction.
• Trains the model end-to-end on a novel sonnet corpus extracted from Project Gutenberg using structural and lexical constraints.
• Uses a character-level language model with attention mechanisms to generate lines with specified stress and rhyme patterns.
• Employs a joint optimization objective that enforces correct iambic pentameter and ABAB/CDCD/EFEF/GG rhyme schemes during generation.
• Leverages pre-trained word embeddings from a background poetry corpus to improve language modeling capacity.
• Applies a multi-stage generation process: first predicting stressed syllables, then generating rhyming word pairs, and finally composing full lines with coherent semantics.

Experimental results

Research questions

• RQ1 Can a joint neural model effectively generate sonnets that maintain correct iambic pentameter and rhyme schemes while preserving poetic coherence?
• RQ2 How do crowdworkers judge the human-likeness of machine-generated sonnets, and which features (e.g., rhyme, meter) do they rely on?
• RQ3 To what extent does expert evaluation reveal limitations in machine-generated poetry that are missed by crowd judgments?
• RQ4 Does explicit modeling of meter and rhyme significantly improve poetic quality compared to a vanilla language model?
• RQ5 Does focusing on formal constraints (meter and rhyme) negatively impact readability and emotional depth in generated poetry?

Key findings

• The stress and rhyme models perform exceptionally well, with generated quatrains rated as nearly indistinguishable from human-written poems by crowdworkers.
• Crowdworkers achieved 53.2% accuracy in identifying human-written poems when using the full model (LM∗∗+PM+RM), indicating high human-likeness in form and style.
• Expert evaluation revealed that the full model scored 4.43/5.0 for rhyme and 4.10/5.0 for meter—slightly higher than human poets—demonstrating strong formal accuracy.
• Despite strong form adherence, the full model scored only 2.70/5.0 for readability and 2.90/5.0 for emotion, significantly lower than human poems (4.80 and 4.37, respectively).
• A vanilla language model (LM) achieved 4.00/5.0 for meter, suggesting it implicitly learns iambic pentameter without explicit modeling, challenging the necessity of dedicated meter modules.
• The model without the meter component (LM∗∗+RM) achieved identical crowdworker accuracy (0.532), indicating that rhyme alone is a dominant cue for layperson judgment, while meter is largely ignored in casual evaluation.