so much depends / upon / a whitespace: Why Whitespace Matters for Poets and LLMs

TL;DR

Whitespace in poetry encodes both semantic and visual meaning, challenging common NLP assumptions about text structure. The authors introduce the Whitespace In Spatial Poetics (WISP) typology and conduct a large-scale, multimodal analysis across 19k published, 12k unpublished, and 51k generated English poems, releasing a 2.8k-poem public-domain subset with preserved formatting. They evaluate multiple linearization pipelines (HTML-to-text, a custom WISP-ify, and OCR-based approaches) and demonstrate that whitespace preservation is highly sensitive to data processing, with implications for pretraining data curation and multimodal model design. The work highlights how whitespace usage varies by form, era, topic, and data source, and argues for incorporating whitespace-aware pipelines (and TEI-inspired encoding) to better preserve authors' intended layouts in future NLP and LLM applications.

Abstract

Whitespace is a critical component of poetic form, reflecting both adherence to standardized forms and rebellion against those forms. Each poem's whitespace distribution reflects the artistic choices of the poet and is an integral semantic and spatial feature of the poem. Yet, despite the popularity of poetry as both a long-standing art form and as a generation task for large language models (LLMs), whitespace has not received sufficient attention from the NLP community. Using a corpus of 19k English-language published poems from Poetry Foundation, we investigate how 4k poets have used whitespace in their works. We release a subset of 2.8k public-domain poems with preserved formatting to facilitate further research in this area. We compare whitespace usage in the published poems to (1) 51k LLM-generated poems, and (2) 12k unpublished poems posted in an online community. We also explore whitespace usage across time periods, poetic forms, and data sources. Additionally, we find that different text processing methods can result in significantly different representations of whitespace in poetry data, motivating us to use these poems and whitespace patterns to discuss implications for the processing strategies used to assemble pretraining datasets for LLMs.

Figures (13)

• Figure 1: An excerpt from "[Buffalo Bill 's]" (1926) by E.E. Cummings, annotated using our whitespace typology, WISP(Whitespace In Spatial Poetics). WISP distinguishes between five categories of whitespace usage: line breaks, prefix space, internal space, vertical space, and line length.
• Figure 2: Prefix whitespace lengths, Published Poems.
• Figure 3: Comparison of prefix and internal mean whitespace usage across the source datasets. To ensure a fair comparison, we compare the generated poems (which are almost all free-verse) only to free-verse poems from Poetry Foundation (as tagged on the website) and Reddit (as predicted using a prompt; see §\ref{['subsection:data-reddit']}).
• Figure 4: Prefix and internal whitespace usages over time. The y-axis shows the mean number of spaces included in the whitespace, for all non-standard whitespace usages (we excluded non-standard usages from the denominator to highlight increasingly bold usages over time). Shaded areas show 95% confidence intervals, and period lines are based on the Norton Anthology of English Literature, 11th edition.
• Figure 5: Comparison of most frequent dependency triples that span line breaks across the source datasets.