[Paper Review] Shall I Compare Thee to a Machine-Written Sonnet? An Approach to Algorithmic Sonnet Generation

John Benhart, Peter Hase+2 more|arXiv (Cornell University)|Nov 13, 2018

Topic Modeling19 references4 citations

TL;DR

This paper presents a hybrid neural and rule-based approach to generating Shakespearean sonnets with precise meter, rhyme, and grammatical structure. By integrating in-line punctuation, part-of-speech constraints, and curated training corpora, the method produces high-quality, human-like poetry that won the 2018 PoetiX Literary Turing Test, outperforming prior systems in formal accuracy and poetic coherence.

ABSTRACT

We provide an approach for generating beautiful poetry. Our sonnet-generation algorithm includes several novel elements that improve over the state of the art, leading to metrical, rhyming poetry with many human-like qualities. These novel elements include in-line punctuation, part of speech restrictions, and more appropriate training corpora. Our work is the winner of the 2018 PoetiX Literary Turing Test Award for computer-generated poetry.

Motivation & Objective

• To develop a system capable of generating Shakespearean sonnets that adhere strictly to iambic pentameter and the ABABCDCDEFEFGG rhyme scheme.
• To improve upon existing neural poetry generation methods by incorporating hard constraints on meter, rhyme, and part-of-speech usage.
• To produce sonnets with human-like qualities such as punctuation, grammatical correctness, and stylistic coherence.
• To evaluate the quality of generated poetry through expert human judgment in a Turing test setting.
• To demonstrate that combining neural language models with domain-specific rules yields more linguistically and poetically accurate results than purely statistical approaches.

Proposed method

• The method uses a neural language model trained on curated corpora of Shakespearean sonnets to generate lines with high likelihood.
• It applies beam search to select lines that maximize likelihood under the model while satisfying hard constraints on meter and rhyme.
• The system uses the CMU Pronouncing Dictionary to enforce iambic pentameter by assigning stress patterns (0 for unstressed, 1 for stressed) to each word.
• It enforces part-of-speech-based sampling constraints to prevent ungrammatical phrases, such as adjacent nouns or comparative adjectives and verbs.
• In-line punctuation is explicitly modeled and generated to improve syntactic clarity and poetic rhythm.
• Rhyme words are selected based on cosine similarity between the user prompt’s GloVe embedding and precomputed rhyme pairs, ensuring thematic and phonetic coherence.

Experimental results

Research questions

• RQ1 Can a hybrid neural and rule-based system generate Shakespearean sonnets that strictly adhere to iambic pentameter and the ABABCDCDEFEFGG rhyme scheme?
• RQ2 How do part-of-speech constraints and in-line punctuation affect the grammatical and stylistic quality of algorithmically generated poetry?
• RQ3 To what extent can curated training corpora improve the stylistic fidelity of generated sonnets compared to general-purpose text?
• RQ4 Can a machine-generated sonnet achieve high enough quality to win a human-judged literary competition like the PoetiX Literary Turing Test?
• RQ5 How do hard constraints on meter and rhyme compare to soft constraints in terms of poetic accuracy and human perception?

Key findings

• The proposed method successfully generates sonnets that consistently follow iambic pentameter and the ABABCDCDEFEFGG rhyme scheme, with no metrical or rhyming errors.
• Part-of-speech constraints effectively eliminate ungrammatical phrases such as 'human beings procreation' or 'greater be', which appeared in prior systems.
• The inclusion of in-line punctuation significantly improves syntactic clarity and poetic fluency compared to systems that omit such features.
• The system outperformed prior state-of-the-art approaches in the 2018 PoetiX Literary Turing Test, winning first place despite being correctly identified as machine-generated by expert judges.
• Human judges, trained in sonnet analysis, were able to distinguish all machine-generated poems from human ones, yet still awarded the system the top prize, indicating high perceived poetic quality.
• The method’s use of hard constraints and curated training data resulted in higher line likelihoods and greater formal precision than systems relying on soft constraints or less relevant corpora.