Source-linked AI summary
Zero-shot Sonnet Generation with Discourse-level Planning and Aesthetics Features
Yufei Tian, Nanyun Peng
TL;DR
Poetry generation often lacks large training datasets, while full-length sonnets must meet meter and rhyme constraints and remain coherent and aesthetic. Zest plans a sonnet sketch using non-poetic texts, adds rhyme and figurative features, and applies constrained decoding; evaluations report more coherent, poetic, and creative sonnets than baselines. The approach does not capture stanza-level structural characteristics during planning and generation.
Problem
The paper targets generating coherent, aesthetically rich full-length sonnets without training on poetic texts, in a setting where poetic training data are scarce.
Source: §1 · p. 1Method
Zest plans a sonnet sketch from non-poetic texts, forms rhyme pairs, polishes keywords with imagery and simile, and uses meter-and-rhyme-constrained decoding.
Source: §1 · p. 2Results
Human evaluation reports that Zest generates more discourse-level coherent, poetic, creative, and emotion-evoking sonnets than baselines.
Source: §1 · p. 2Takeaways & Limitations
The conclusion reports that the approach generates globally coherent, poetic, emotion-provoking sonnets with rich imagery without training on poems.
Source: §8 · p. 9The current approach does not capture stanza-level structural characteristics during planning and generation.
Source: §6.3 · p. 8
Abstract
from arXiv · showhide
Poetry generation, and creative language generation in general, usually suffers from the lack of large training data. In this paper, we present a novel framework to generate sonnets that does not require training on poems. We design a hierarchical framework which plans the poem sketch before decoding. Specifically, a content planning module is trained on non-poetic texts to obtain discourse-level coherence; then a rhyme module generates rhyme words and a polishing module introduces imagery and similes for aesthetics purposes. Finally, we design a constrained decoding algorithm to impose the meter-and-rhyme constraint of the generated sonnets. Automatic and human evaluation show that our multi-stage approach without training on poem corpora generates more coherent, poetic, and creative sonnets than several strong baselines.
1 Introduction
Zest addresses the scarcity of poetic training data and the challenge of generating coherent, aesthetically rich full-length sonnets. It separates planning from decoding, adds poetic features to a sonnet sketch, and uses constrained decoding; human evaluation reports improvements over baselines.
1 Introduction
- Only 3,355 sonnets were collected in the cited resource, while training on related creative corpora may be infeasible where poetry or lyrics data are scarce.The authors also argue that learning from and mimicking existing corpora is unlikely to produce novel content.Source: §1 · p. 1
- Prior work often prioritized meter and rhyme or short stanzas; full-length sonnets remained a discourse-coherence challenge without translation from another language.The passage describes topical rhyme words as insufficient for discourse-level coherence and notes a translation-based approach.Source: §1 · pp. 1–2
- Zest plans a sonnet sketch using non-poetic texts, forms rhyme pairs, polishes keywords with imagery and simile, then decodes under meter-and-rhyme constraints.The content-planning and final-decoding components are trained on news and story corpora; the polishing step uses external knowledge.Source: §1 · p. 2
- The authors introduce automatic format checking and novelty evaluation, including diversity and imageability, as measures of generated-poem quality.Source: §1 · p. 2
- Human evaluation finds that Zest generates more discourse-level coherent, poetic, creative, and emotion-evoking sonnets than baselines.Source: §1 · p. 2
2 Background
The background introduces sonnet structure, meter, rhyme, and the terminology used to describe the poem sketch.
2 Background
- The section introduces sonnet structure, meter, rhyme, and key terminology for the generation task.Source: §2 · p. 2
2.1 The Structures of Sonnets
- Shakespearean sonnets use three quatrains and a rhyming couplet, while Petrarchan sonnets use an octave followed by a sestet.The Shakespearean example rhyme scheme is ABABCDCDEFEFGG.Source: §2.1 · p. 2, Table 1 · p. 2
2.2 Meter Constraints
- Iambic pentameter has ten alternating unstressed and stressed syllables, with a reversed first foot and an added final unstressed syllable among described variations.Source: §2.2 · pp. 2–3
2.3 Rhyme Words, Couplets and Patterns
- Rhyme pairs share the same or similar ending sounds; slant rhymes share stressed vowels but have similar, not identical, ending consonants.Source: §2.3 · p. 3
2.4 Terminology
- A poem sketch comprises content words, rhyme words at line endings, and keyword modifications for aesthetics; keywords combine content and rhyme words.Source: §2.4 · p. 3
3 Approach
Zest builds a sonnet sketch before decoding, using content planning, rhyme pairing, aesthetic polishing, and constrained final decoding. Its modules use non-poetic text and external resources to plan keywords, add figurative devices, and generate metrical, rhyming lines.
3 Approach
- The four-stage framework plans content, forms rhyme pairs, adds imagery and similes, then decodes the sketch under meter-and-rhyme constraints.The first three stages provide salient points for the sonnet sketch; final decoding turns it into a well-formed sonnet.Source: §3 · p. 3
3.1 Content Planning
- T5 predicts body keywords from a title, using extracted salient words from news and stories as training outlines.One [MASK] placeholder corresponds to one output word, allowing control of the number of lines and keywords at inference.Source: §3.1 · p. 3
3.2 Generating Rhyme Words
- The system selects an initial rhyme word from each initial rhyming line, then retrieves strict and slant rhyme candidates using phonetic information.The outline contains two or three content words per line, with the last keyword in initial rhyming lines selected as the initial rhyme word.Source: §3.2 · p. 4
3.3 Polishing Context Words for Aesthetics
- Polishing adds imagery from ConceptNet-based pairs and simile vehicles generated from sketch adjectives, retaining the most confident candidates.Simile phrases are also filtered to conform to iambic meter.Source: §3.3 · p. 4
3.4 Sketch to Sonnet Generation
- GPT-Neo-2.7B generates the sonnet line by line in reverse, incorporating sketch keywords and prior lines while filtering tokens to enforce meter.Simile phrases are fixed during generation, and decoding continues until 10 or 11 syllables form a line with all context words incorporated.Source: §3.4 · p. 5
4 Experimental Setup
The experiments train the planning and decoding modules on news and story data, compare Zest with existing and ablated systems, and evaluate format, novelty, and expert judgments. Human ratings use six literature-trained professionals, with three raters per sonnet.
4.1 Dataset
- Training uses 4,500 CNN news summaries and 16,000 Reddit short stories, excluding pieces outside the specified length range or containing conversations or URLs.T5 is fine-tuned for 10 epochs and GPT-Neo-2.7B for 6 epochs.Source: §4.1 · p. 5
4.2 Baselines
- Baselines include lyrics-trained Hafez and few-shot GPT-3, while Prosaic and Zest without figurative polishing test planning and polishing variants.Prosaic is adapted to GPT-2 for comparison with Zest.Source: §4.2 · p. 5
4.3 Decoding Strategy
- Decoding uses top-k sampling and a three-token no-repeat constraint to encourage creativity and avoid repetition.The model continues generation until it completes a word, enabling meter and rhyme constraints to be applied.Source: §4.3 · pp. 5–6
4.4 Automatic Evaluation
- Automatic evaluation measures rhyme, meter, and syllable validity, plus diversity with Distinct-2 and imageability from word ratings.Table 2 summarizes results for rhyme, meter, syllable checking, distinct scores, and imageability.Source: §4.4 · p. 6, Table 2 · p. 6
4.5 Human Expert Judgement
- Six literature-trained professionals rate 50 sonnets per baseline and the human gold standard, with each sonnet judged by three professionals.Pearson inter-annotator agreement is 0.61 (p-value <0.01), and the best model differs significantly from the best baseline in a paired t-test.Source: §4.5 · p. 6
5 Results and Analysis
Zest performs strongly on automatic and expert-rated sonnet criteria, with significant human-rating gains over the best baseline. Its full version trades some coherence for higher originality and poeticness, while the paper attributes its poeticness to imagery, concise keyword planning, and constrained decoding.
5.1 Results of Automatic Evaluation
- Human poets vary more and follow constraints less strictly, while GPT-3 fails to learn sonnet formats and Prosaic falls short on meter-checking.Source: §5.1 · p. 6
5.2 Results of Human Evaluation
- Zest and Zest w/o fig significantly outperform the best baseline across coherence, originality, poeticness, and emotion evoking (p < 0.05).Table 3 reports expert ratings on a 1–5 scale.Source: Table 3 · p. 7, §5.2 · p. 7
5.1 Results of Automatic Evaluation
- Zest pairs rhyme words with 100% success and is the most diverse and best at imageability among machine systems.Source: §5.1 · p. 7
5.2 Results of Human Evaluation
- Against Prosaic, both Zest models produce significantly more coherent sonnets (p-value < 0.01), supporting explicit sketch planning over generation from scratch.Source: §5.2 · p. 7
- Zest w/o fig is more coherent than Zest (p-value < 0.10), while Zest scores +0.2 higher in originality and poeticness and leads machines in emotion evoking.Source: §5.2 · p. 7
- Zest is on par with humans in poeticness; the authors link this to imagery and similes, at least three planned concepts per line, and occasional omissions of function words.The decoding constraint can make language less natural, but the authors say occasional omissions can suit sonnets.Source: §5.2 · p. 7
6 Qualitative Analysis
The case study contrasts disconnected topical rhyme and prosaic language in baselines with Zest’s organized seasonal progression and vivid imagery. Keyword guidance succeeds in most evaluated cases, while the approach does not capture stanza-level structure and is only proposed for extension to other languages.
6.1 Case Study
- Hafez’s topical rhymes do not ensure overall coherence, and Prosaic produces colloquial, prosaic sentences; Zest w/o fig orders the seasons logically while full Zest adds vivid imagery.The example describes the four seasons from winter to autumn and notes minor grammar errors in Zest.Source: §6.1 · p. 7
6.2 Impact of Keywords
- Planned keywords successfully guide generation 82% of the time; reported failure causes include novel unplanned content (8%) and polishing that alters keyword meaning (6%).Source: Figure 4 · p. 8
6.3 Limitation and Future Direction
- The approach does not capture stanza-level structure, where lines interlace more within a stanza and a new stanza usually signals a viewpoint transition.Source: §6.3 · p. 8
- The authors plan to extend the pipeline to other languages and state that pretrained language models and existing rhyme resources make its first two components transferable.Source: §6.3 · p. 8
7 Related Work
Prior poetry-generation work often targets short stanzas or uses topical rhyme words, while coherent sonnets without cross-language translation remain challenging. Zest instead separates content planning from decoding and explores figurative language together with meter-and-rhyme constraints without poetry data.
Poetry Generation
- Several neural poetry systems target only a first four-line stanza (Wang et al., 2016; Lau et al., 2018; Yi et al., 2018; Liu et al., 2019).Source: §7 · pp. 8–9
- The authors describe their work as the first to combine figurative speech and meter-and-rhyme constraints in poetry generation without poetry data.Source: §7 · p. 9
Content Planning
- Earlier plan-and-write systems learn plots from the target domain, whereas Zest separates planning from decoding to address the shortage of sizable creative-text training data.Source: §7 · p. 9
8 Conclusion
The paper presents a poem-free, hierarchical sonnet-generation framework and reports that it produces globally coherent, poetic, emotion-provoking sonnets.
8 Conclusion
- The model generates sonnets without training on poems, using high-level content planning, poetic-feature refinement, and decoding-time meter-and-rhyme control.Source: §8 · p. 9
- Its hierarchical framework plans high-level content, refines keywords with poetic features, and imposes meter-and-rhyme constraints during decoding.Source: §8 · p. 9
- Extensive automatic and expert evaluation finds the generated sonnets globally coherent, poetic, and emotion provoking, with rich imagery.Source: §8 · p. 9