Abstract
This research introduces a quantitative framework to map the stylistic "fingerprint" of American poetry into a high-dimensional vector space. By analyzing orthographic, syntactic, and phonemic features rather than just semantic content, the authors demonstrate that structural and sonic patterns are significantly more effective than traditional word-frequency methods for identifying a poet’s unique voice. Using Principal Component Analysis (PCA), the study visualizes these stylistic relationships, revealing distinct clusters that align with qualitative literary analysis.
Addressing the "How" of Poetry
Most Natural Language Processing (NLP) focuses on semantics—extracting the "what" from a text. While effective for informational prose, this approach fails to capture the essence of poetry, where the "how" (style) is often paramount. Traditional text analysis relies on "bag-of-words" models like TF-IDF, which treat text as a disorganized collection of words, missing the rhythmic, structural, and phonetic nuances that define a poet’s voice. This study bridges the gap by treating style as a multi-layered latent structure, allowing for objective comparison and visualization.
Methodology: The 84-Dimensional Style Vector
The authors developed a feature extraction engine that maps every poem to an 84-dimensional vector across three domains:
- Orthographic (Structure): Metrics such as word count, line length, and stanza count.
- Syntactic (Grammar): Rule-based Part-of-Speech (POS) tagging to analyze discourse modes, such as the ratio of adjectives to verbs.
- Phonemic (Sound): Utilizing the CMU Pronouncing Dictionary to identify rhyme types (identity, perfect, slant), alliteration, assonance, and consonance.
To make this data interpretable, Principal Component Analysis (PCA) projects these high-dimensional vectors onto a 2D plane, where physical proximity represents stylistic similarity.
Datasets and Scalability
The primary analysis utilized a curated collection of 81 poems by 18 prominent American poets (e.g., Robert Frost, Sylvia Plath, Walt Whitman) sourced from The Oxford Anthology of Modern American Poetry. The authors also tested the scalability of their PCA visualization on samples of up to 100,000 poems.
Key Insights
- Style vs. Topic: Unlike TF-IDF, which clusters poems by subject matter (e.g., "bees"), the stylistic model groups poets by structural habits regardless of theme.
- Literary Mapping: Visualization placed Frank O’Hara between the traditionalism of Frost and the modernism of Moore, providing a quantitative basis for existing literary criticism.
- Consistency: While individual short poems vary, a poet’s collective body of work tends to gravitate toward a consistent stylistic "center of gravity."
Quantitative Results
The "Style" algorithm consistently outperformed traditional TF-IDF in differentiating individual voices. In the visualization of the Oxford Anthology, poems by the same author formed distinct, recognizable clusters.
As shown in the comparative analysis, the stylistic features provided a much higher normalized difference between inter-poet and intra-poet distances compared to frequency-based methods (TF and TF-IDF), particularly in the diverse Oxford dataset.
Limitations and Future Directions
The current model does not yet account for prosody (rhythmic variations in stress) or visual style (typography and spacing, as seen in E.E. Cummings). Future work may explore more complex dimensionality reduction techniques like Non-negative Matrix Factorization (NMF) to capture the evolution of a poet's style over time.
Summary updated