Mining poetic rhythm: using text-to-speech software to rewrite English literary history

1.1. Overview
Automatically predicting where the beats fall in a line of English poetry is difficult. This is because the stress placed on a word will be dependent upon the meaning of the statement of which it is a part. Without approximating the meaning of a line, then, it is not possible to arrive at an accurate prediction as to where the beats will fall when the poem is read or performed. This is no trivial task. Yet without determining the beats, it is impossible to identify the metre of the line or its sound-patterning, such as alliteration and assonance, which are instrumental to the rhythm. Because of this difficulty, the history of English poetic rhythm remains almost entirely uncharted. Twenty years ago, Preminger and Brogan's New Princeton Encyclopaedia of Poetry and Poetics lamented that 'there is at present no comprehensive and reliable history of the development of metrical practice in the West, not even competent accounts for any one language'.1 The intervening decades have not resolved this problem for English. While there have been notable attempts to construct a reliable history, none have been comprehensive. Working without the benefit of computer automation, literary historians have focused exclusively on a small number of major poetic figures.

Having done so, the current histories exhibit erroneous assertions of innovations in rhythm and rhyme, give credit to the wrong poets, and neglect formal experiments occurring on the margins. Further, most approaches tend to rely on the impressions of the critic, rather than empirical data. Consequently, the majority of statements about poetic rhythm have been made in a position of ignorance about the formal characteristics of the corpus of English poetry.

The problem becomes acute from the start of the nineteenth century as the volume of printed material explodes. Nevertheless, Martin J. Duffell's New History of English Metre (2008) considers only 32 major poets for the period, analysing 200-300 lines of poetry as representative samples of their practice.2 His analysis of poetic form therefore encompasses fewer than 9,000 lines. This is grossly inadequate, given that W. B. Yeats's Collected Poems alone surpasses this total by 2,000 lines. The number of poems considered by Duffell is a tiny proportion of the 151,299 poems in English considered important enough to be included in the Literature Online (LION) database for this period.

1.2. Methodology
The problem of coverage can be overcome by harnessing the power of 21st Century achievements in speech synthesis and text-to-speech software. My three-year project at King's College London, funded by the British Academy, is concerned with doing just this. It uses the MARY text-to-speech software, developed by the DFKI (the German Research Centre for Artificial Intelligence) and the Institute of Phonetics at Saarland University. The intended outcome of the project is a more reliable and comprehensive account of developments and trends in poetic rhythm, metrical forms, sound-patterning, rhyme schemes, and stanza types, in verse in English for the period of 1800--1970.

The MARY-tts software has recently been used to visualize sound patterns in literary texts by Tanya Clement et al, re-presenting aural data as high-lighted text.3 My approach differs by extracting key data from an intermediate stage of natural language processing that enable the software to identify some of the beats in a line of poetry. My scripts then extrapolate where the other beats are likely to fall, if the line be read as poetry rather than normal speech. From these data, my tools can distinguish between binary and ternary metres, determine the number of beats per line in the metrical template, and identify trochaic and iambic verse. It can spot refrains, and repeated structures, and sound-patterning such as internal rhyme and alliteration. The approach has been tested on verse selected from a wide variety of sources, including Wordsworth, Browning, Keats, Tennyson, Hopkins, Eliot, and the entire corpus of Yeats's poems.

There have been earlier attempts to automatically determine formal qualities of poems using digital tools. Marc Plamondon's tool developed for the Representative Poetry Online database at Toronto remains an impressive example.4 However, none of these approaches have employed text-to-speech software to predict phonemic properties of the text automatically. The manual entry of syllables and phonemes for each word inevitably limits the scope and accuracy of such tools. Other impressive attempts to uncover the grammatical rules of particular poets' prosody have been similarly limited by manual input of language processing.5

1.3. Conclusion
My preliminary results suggest that the project will succeed in its aim of massively enlarging and enriching our understanding of literary history. Further, in assigning credit to the real progenitors of formal trends, and in providing the data with which to analyse formal developments empirically, rather than impressionistically, this project has the potential to rewrite the history of poetic form, and of poetic influence, altogether.

References
1. Preminger, Alex and Brogan, T. V. F and Warnke, Frank J. and Hardison, Jr., O. B. and Miner, Earl.The New Princeton Encyclopedia of Poetry and Poetics. Princeton University Press, Princeton, N. J.: 1993. p. 777.

2. Duffell, Martin J. (2008), A New History of English Metre Modern Humanities Research Association and Maney Publishing, London.

3. Plamondon, M. R. (2006). Virtual Verse Analysis: Analysing Patterns in Poetry. Literary and Linguistic Computing. 21. Supplemental Issue: 127-141.

4. Hayes, Bruce and Wilson, Colin and Shisko, Anne (2012). Maxent grammars for the metrics of Shakespeare and Milton. Language 88:4 (December 2012): 691-731.