The Stanford Literary Lab Transhistorical Poetry Project Phase II: Metrical Form

To date, most studies that foreground quantitative analyses of literature have focused exclusively on prose writing (the novel in particular) rather than poetry (Stanford Literary Lab 2011, Clement 2008). In part, this state of affairs is due to poetry’s highly figurative language and complex communicative intent, which poses acute problems for text mining and similar quantitative analyses that rely on lexical and semantic meaning and whose methodological origins lie in the “hard” sciences (Pasanek and Sculley 2008, Bei 2008). Poetry, however, offers a unique subject for quantitative analysis, independent of lexical and semantic meaning, that is largely absent from prose works: meter. The practice of scansion is an ancient study that precedes the advent of the English language. It has, nevertheless, always consisted mainly of counting, sorting, and indexing words and word components, endeavors to which quantitative analysis is especially attuned. Despite this apparent sympathy between metrics and quantitative analysis, however, the algorithmic detection of the complexities of meter remains outside of the current capabilities of the Digital Humanities. The irregularity of stress, syllabic schemes, and the rule-bending nature of poetic diction runs counter to the binary presence/absence process of most computational analysis.

In the second phase of the Stanford Literary Lab’s multi-year ongoing project to create a system for detecting the formal features of poetry, we have focused our attention on the question of meter. Using a new method that combines a series of rule-based analyses with an iterative probabilistic-based classification algorithm, we can now detect, with a high degree of accuracy, both the meter and line length of individual poems. We have trained our algorithm to recognize individual metrical feet, such as iambs, dactyls, anapests, trochees, and spondees, and to combine these identifications with a signal-processing approach to the entire poem to classify the overall metrical scheme of any given poem. We have trained and tested our algorithm on a corpus of over 300,000 English language poems from the late medieval period to the twentieth century. Moreover, we have also applied our algorithm to multi-lingual corpora: in our presentation we will demonstrate how our methods are effective on German and French, as well as English poetic forms.

This project builds upon the first phase of our project, presented to the Alliance of Digital Humanities Organizations conference in Lincoln, Nebraska, which successfully sought to recognize the syllabic scheme patterns in poetic lines. The overarching goal of our project as a whole is perhaps simple to state, but challenging to execute: we seek to create a program that can automatically identify poetic forms. In other words, we are in the process of designing a program that can read any number of poems and tell us their exact syllable count, meter, rhyme scheme, use of traditional forms (e.g., sonnet, ballad, sestina), etc. Success in creating such a program would represent an important tool for scholars in the Digital Humanities, and would offer the ability to:

1) Create an inventory of all poetic forms, traditional and untraditional.

2) Trace the history of poetic forms, including:

- variation among poetic forms (e.g., Which forms were most popular in a given period? Were some periods more formally diverse than others? How does form diversity change over time?)

- variation within poetic forms (e.g., How do the forms themselves change over time? Are sonnets more metrically rigorous in one era than another?)

3) Better understand the relationship between form and meaning by relating analyses of scansion with those of lexical and semantic meaning.

4) Provide distant readings to help generate and/or support new close readings.

In constructing such a program, we break down the task of recognizing metrical schemes into the simpler, but by no means simple, components of recognizing: 1) Syllable Scheme; 2) Beat Scheme; 3) Rhyme Scheme; 4) Metrical Scheme; and then 5) matching any combination of these categories to a tradition name (e.g., sonnet, heroic couplets, rhyme royale, etc.) if one exists to describe it.

We began by designing a program that could accurately detect the number of syllables in a given line of poetry (item 1 above) because we believed it would be the most straightforward element to analyze. We additionally realized that if we limited our sample according to metrical foot, then syllable count would present a shortcut toward detecting a rough approximation of meter. For example, if we started by analyzing only iambic poems (as recognized by human readers), and our syllabifier counted 9-11 syllables in each line of a given poem, then we could have reasonable assurance that the poem was written in pentameter. A similar process could be applied to other metrical patterns. In training our syllabifier, we purposely limited our sample to poems composed from roughly the late sixteenth century to the late nineteenth century because metrical forms were most stable and recognizable in this period.

Building upon this earlier success, we have combined our ability to recognize syllabic scheme with a complex approach to meter that has, at present, an over 80% success rate in correctly classifying meter. Our presentation will discuss how our algorithm was built, the specific challenges that we faced (e.g., elision, extrametrical syllables, feminine endings, foreign words, and other features of meter that are commonly acceptable in the practice of poetics but that our program had difficulty overcoming), and we will present the results of our application of this technique to a large, multi-lingual corpus that shows the historical shape of various metrical forms important to European poetry from the sixteenth to the late nineteenth century. Our initial analysis, using only syllabic scheme, revealed a number of significant and unexpected observations, including the fact that the use of pentameter peaked around the middle of the seventeenth century and has been on the decline ever since, as well as the fact that the use of tetrameter has been reciprocally on the rise since the early eighteenth century and is today equally as popular as pentameter. In this second phase, we expand the results of this analysis to show the historical prominence of the sonnet and heroic couplet forms, the transnational inheritance of metrical form and the history of iambic pentameter in English poetry.

We believe that what we have achieved with this work so far will aid in future quantitative and digital work on poetry, a lacuna that represents a critical problem for the use of digital humanities in the study of literature, given the enormous significance of the poetic tradition within literary studies. We also believe, moreover, that our program has application well beyond the study of poetry and could help to analyze and detect metrical schemes in song, drama, and prose as well, generating topics of analysis that likely would remain undetected without quantitative analysis.