TL;DR
This paper investigates how poetry and prose have evolved and converged over the last 150 years, particularly in the age of mechanical reproduction. By analyzing stylistic features like grammar and meter, the authors demonstrate a significant blurring of lines between poetry and prose in recent times, suggesting a re-evaluation of traditional definitions.
What Problem Does This Paper Address?
The paper tackles the challenge of defining poetry in an era where mechanical reproduction and AI have democratized content creation. Traditional definitions, often rooted in esoteric philosophical concepts, struggle to keep pace with the evolving landscape of literary expression. Specifically, the authors note that semantic similarities between prose and poetry make distinguishing them difficult, and they aim to explore how stylistic features have changed over time to understand this convergence. They question whether the historical distinctions between poetry and prose have diminished, particularly in the context of modern digital media.
Methods
The authors employ a computational approach to analyze historical texts. They focus on stylistic features rather than semantic content, hypothesizing that these are more robust differentiators. Key features include:
- Grammar: This encompasses various types of inversions (e.g., subject-verb inversion, prepositional phrase inversion, Yoda construction), dependent clauses as subjects, rhetorical questions, and lines beginning with conjunctions. These are identified using natural language processing tools like Stanford CoreNLP for tokenization, part-of-speech tagging, and dependency parsing.
- Meter: They analyze the rhythmic structure of text by identifying base meters (iambic, trochaic, anapestic, dactylic) and their feet using the Stanford Literary Lab's Poesy module.
- Rhyme: While primarily used for comparing poetry across different periods, rhyme schemes were also extracted.
For classification, they trained Random Forest and KNN classifiers on feature vectors derived from these stylistic elements. They also addressed class imbalance by adjusting weights inversely proportional to class frequencies.
Datasets
The study utilizes four datasets, curated from two distinct time periods:
- 1870-1920: This period represents the earlier stages of mechanical reproduction and the flourishing of romantic poetry.
- 1970-2019: This later period reflects the digital age and widespread use of artificial intelligence.
For poetry, content was sourced from PoetryFoundation and PoemHunter websites, focusing on works by popular poets of each era. For prose, they selected "top liked quotes" from popular books listed on Goodreads, reasoning that these quotes often exhibit poetic qualities and are contextually independent.
Key Insights and Ideas
- Semantic vs. Stylistic Distinction: The paper strongly argues that semantic content alone is insufficient to distinguish poetry from prose, as many prose excerpts can be highly poetic. Stylistic features, particularly grammatical ones, offer a more reliable differentiation.
- Convergence of Poetry and Prose: The most striking insight is the significant convergence between poetry and prose in the 1970-2019 period compared to the 1870-1920 period. This is evidenced by a decrease in classification accuracy between poetry and prose in the later period.
- Decline of Traditional Poetic Features: Traditional poetic features like meter and rhyme appear to be less prevalent in contemporary poetry, with a notable increase in "no meter" or "no rhyme scheme" classifications.
- Inversion as a Key Differentiator: Grammatical inversions, often used to fit poetic lines into specific meters or rhyme schemes, are identified as a crucial stylistic feature that has historically differentiated poetry from prose. The study shows a decline in inversion counts in modern poetry.
- Prose Remains Relatively Stable: In contrast to poetry, the stylistic features analyzed in prose show minimal change between the two historical periods, suggesting prose has maintained a more consistent structure.
Key Results
- Classification Accuracy:
- Poetry vs. Prose (1870-1920): Achieved high accuracy (e.g., 98.4% with all features using Random Forest).
- Poetry vs. Prose (1970-2019): Accuracy significantly decreased (e.g., 91.5% with all features using Random Forest), indicating increased similarity.
- Poetry (1870-1920) vs. Poetry (1970-2019): Achieved 77% classification accuracy, highlighting significant changes in poetry over time.
- Prose (1870-1920) vs. Prose (1970-2019): Showed very low classification accuracy (around 59%), suggesting minimal change in prose.
- Overall Poetry vs. Prose (combined periods): Achieved high accuracy (94.7%), demonstrating that even with convergence, distinct features still allow for classification.
- Meter and Rhyme Trends: Figures show a decline in the dominance of iambic meter and a significant increase in poetry with "no recognizable meter" in the 1970-2019 period. Similarly, a large percentage of contemporary poetry lacks a rhyme scheme.
Limitations and Future Work
Limitations:
- The study acknowledges that "quotes" were used as a proxy for prose, which might not fully represent all prose characteristics.
- Line breaks, a traditional visual differentiator, were intentionally excluded to focus on deeper stylistic features.
- The analysis of rhyme was limited to comparing poetry across periods, not between poetry and prose.
Future Work:
The authors propose developing a personalized poetry assistant. This assistant would learn a user's stylistic preferences in terms of inversions and meter, allowing them to create text that aligns with either "modern" or "classic" poetic styles. This suggests a future direction in leveraging these identified stylistic features for creative AI applications.
Summary updated