Poetry is a special form of literature that employs language in unusual ways to evoke emotion, paint vivid imagery, or convey a message in musical and imaginative ways, and it can draw intense emotion out of readers in manners that regular prose cannot (Abdul-Rahman et al., 2013). To do so, imagery, metaphor, rhyme, meter, and other literary devices are employed to convey meaning and aesthetics in pieces of structured or semi-structured text named stanzas, which are combined into large pieces known as poems. Poetry contains a high density of creativity and ambiguity and it encompasses a wide range of forms, styles, and themes, such as politics, love, or the natural world, to name a few.

In the field of computer-based automated poetry analysis, a method focus on stylometry, which involves measuring statistical features such as the frequency of words or syllables or entropy. This is done to compare styles across groups of works from different cultural movements, periods, or places, in order to uncover recurrent patterns, uses of language, or expressions that can support a critical analysis of the works. The process of inquiry in stylometry follows a typical data analysis workflow, very similar to the traditional data analysis process of experimental sciences, in which data is collected, pre-processed, analyzed, and finally interpreted (Hullman and Gelman, 2021). This process typically starts with data aggregation according to a set of features of interest, followed by computing statistics of interest for each group, and finally establishing a comparison between the statistics of different groups to identify significant similarities and differences that were found, such as the recurrent use of certain words in this case. Among the typical features employed in stylometric analysis of poetry, we can find verse and poem length, rhyme, use of words and parts of speech, or semantics. However, the analysis of sonic and rhythmical patterns is exclusive to poetry, as it cannot be found in any other kind of literary work.

Poetry scansion refers to the analysis of the rhythmic structure of a poem and involves recognizing and counting specific patterns in a poem, such as the stressed and unstressed syllables that can be found (Agirrezabal, 2017; Navarro-Colorado, 2018). Although scansion is intimately related to syllabification through the prosodic features of the language in question, these two terms should not be conflated, as syllabification refers to the mere splitting of words into their constituent units, the syllables, according to the general rules of a language. Thus, a successful scansion typically depends on the accurate identification of prosodic features, such as stress and lexical tone, which are first obtained through the syllabification of the poem's verses (Navarro-Colorado, 2015; Agirrezabal et al., 2017). On its own, scansion is a useful technique for determining the form and style of poetry from different eras and cultures, and it is widely accepted as a keystone scholarly practice in the field of literary analysis. However, the process of scansion is rather complex and has traditionally been done manually in a highly time-consuming manner. Luckily, recent breakthroughs in automatic text processing have resulted in a surge of language-specific computer-based scansion methods (Agirrezabal, 2017; Navarro-Colorado, 2018; de la Rosa et al., 2020), opening up new exploration avenues for digital poetry corpora in many of the most popular languages of the world (Agirrezabal et al., 2016).

This paper presents a conceptualization of the under-explored (Benito-Santos and Therón Sánchez, 2020) visualization design space of distant and close reading¹ of poetical rhythm, which is tested on a corpus of sonnets by the Spanish poet Federico García-Lorca (1898-1936). Spanish poetry, due to the somewhat rigid accentuation rules of the language “meaning that graphemes and stress can be directly derived from written words using a simple set of rules”, historically has been a good candidate to exemplify scansion algorithms (Agirrezabal, 2017; de la Rosa et al., 2020).

Due to the unavailability of computational methods to produce an automatic scansion of the poems, previous approaches focused on the close reading of the rhythm in manually annotated poems (Abdul-Rahman et al., 2013; McCurdy et al., 2016), or on a distant reading of other features different from rhythm (e.g., semantic) found in a collection of poems (Madnani, 2005; Delmonte, 2015). Thus, ours is the first attempt to produce a visualization scheme that can be scaled to visualize the rhythm of one or many poems employing a completely automatic approach. This was facilitated by our active involvement in trying to make text a “first-class attribute” (Brath, 2018, 2020) in our proposed visualizations, meaning that we recurrently exploit some of its visual attributes to encode data features. This is done through a series of prototypical designs based on the heatmap visualization whose aim is to provide a first approach to the problem of visualizing poetic rhythm, and to drive future user studies that are more aligned with the reality of the problem.

To test our designs, we selected a complete poetic corpus of the Spanish poet, playwright and theater director Federico Garc-a-Lorca², which we compiled from the website https://federicogarcialorca.net/. The corpus contains 225 poems (11,268 verses) from 7 different books and compilations, encompassing García-Lorca's main poetic works from the 1920s until his death in 1936 (see Figure 1).

We employed the Python library Rantanplan (de la Rosa et al., 2020)³ to obtain an automatic scansion of all the poems in the corpus. The library employs a fast, rule-based method to syllabify the words in a poem, and, using several heuristics, tries to adjust the verse to known rhythmical structures of Spanish poetry. In their paper, the authors claim that it obtains accuracy scores of 99.99% on the syllabification task and 92.75% on the scansion task in a corpus of 10,000 annotated verses of sonnets “a sonnet contains 14 hendecasyllable verses” from the Spanish Golden Age (1492–1659). However, the authors also report an important decay in performance when analyzing a collection of 4,300 mixed-meter verses (65.02%). Despite our corpus holds poems of very different structure, we considered the library's performance good enough to support a first exploration of our corpus, an endeavor that has already unveiled some potential lines of enhancement for Rantanplan.

After successful processing of a poem, Rantanplan outputs a list of strings “one per verse” of plus “+” and minus “−” symbols, denoting strong “stressed” and weak “unstressed” syllables (e.g., “–+-+-+-”). These sequences, along with other metadata such as the poem's title, or the collection it belongs to, were stored in JSON format⁴ and visualized in the statistical visualization library Altair (VanderPlas et al., 2018), which was employed to build all designs presented in this paper.

In this study, we set out to characterize the visualization design space of poetic rhythm through a series of exploratory designs. These initial experiments hint at the exciting research opportunities that emerge from the fusion of automatic poetry analysis and visualization, a highly underexplored area that we believe deserves more attention from the visualization community. Despite these initial insights, the presented visualizations also show important shortcomings related to the loss of information that occurs during the aggregation process. In particular, the visualization does not inform about the particular distribution of values on each syllable, hiding much of the variability and richness that lies in the data, a practice that many authors in visualization and the humanities have argued against in the past (Drucker, 2011; Coles, 2016; Therón Sánchez et al., 2019; Windhager et al., 2019). As a consequence, it is difficult to infer confidence intervals around the depicted statistic (i.e., the proportion of stressed syllables). In this paper, we added text (as seen in Figure 4) to overcome this issue but this solution does not take into account the empirical distribution of values around the mean, obscuring important information that is known to be error-prone (Hullman and Gelman, 2021). Thus, new designs that take into account these considerations are necessary to maximize the robustness of the proposed analysis workflows that motivate realistic critical analyses of literary corpora.

In addition to the visualization of uncertainty, during the development of the prototype visualizations, we noticed several other research opportunities that we believe are worth exploring, such as the integration of sonification techniques (Rönnberg, 2021; Aigner et al., 2022) that could help users grasp the poems' rhythmical patterns more efficiently by motivating an augmented reading of the poems in the corpus, with numerous potential benefits (e.g., for didactic purposes, or the creation of derived works).

Beyond these considerations, our prototypes also fail at capturing the important role combinatorics play in the composition of poetic literature (Poibeau et al., 2020). To illustrate this concept, take the example of the accentuation rules of the sonnet. Constraining the placement of the stress at the penultimate syllable leaves the writer with considerably fewer options (word combinations) to create a verse that complies with the said constraint. The inclusion of this kind of elements in the visualization could lead to “what-if” types of pattern analysis in which other possible alternative versions of a poem, employing other word combinations and rhythmical patterns, are explored (Wang et al., 2017). This, combined with the incorporation of elements related to rhyme or semantics, could have important implications in the field of literary analysis, enhancing a wide range of scholarly tasks ranging from traditional critical practices to creative writing and composition. To validate these hypotheses, we are organizing a series of workshops with experts in literary studies within the context of the European project CLS Infra (https://clsinfra.io/), from which we expect to recruit a large user base for driving the studies to come. In this regard, we want to extend an open invitation to all members of the visualization and humanities communities interested in contributing to our research to join us in this exciting project.

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found in the article/supplementary material.

The primary intellectual and creative contributions to this work were made by the lead author, who also drafted the manuscript. The additional authors provided valuable input through their proofreading efforts, and contributed crucial data. All authors reviewed and approved the final version for publication.