Poetry In Motion: Quantified Self Data And Automated Poetry Generation

Eververse is a project which synthesises perspectives from disciplines in the humanities and sciences to develop critical and creative explorations of poetry and poetic identity in the digital age. Deploying tools and methods from poetic theory, data analysis, and Natural Language Generation (NLG): the automatic production of natural language output from a non-linguistic data source (Reiter and Dale, 2000).
Eververse uses data from quantified self (QS) devices to automatically generate and publish poetry which correlates to the wearer/poet’s varying physical states.

Context

One of the more common ideas to be found in different statements and theories of poetry presents the poet as a creative vessel or conduit, admitting the sensory input of the world into their bodies and minds, and producing poetic output in turn. For Walt Whitman, every atom of being and sensation was an appropriate inspiration for poetry, and his works exhaustively catalogued the varieties of personal and public experience in nineteenth-century America. In poetics such as this, art increasingly collapses into the being and identity of the artist. Poet W. B. Yeats famously articulated this conundrum, asking “How can we know the dancer from the dance?” The production processes of print, which intervened between the poet and their desire for immediate and spontaneous expression, are circumvented in the digital age, as poets use the internet and social media to create a poetry that is “vast, instantaneous, horizontal, globally distributed, paper thin, and, ultimately, disposable” (Goldsmith, 2016: 195). The internet is shaping contemporary poetry with these characteristic forms, and by incorporating the modalities of images, gifs, video, and sound.
Eververse seeks to explore networked technologies and their affordances to articulate new and novel means of being a poet in the digital age.

Project description

Eververse sends biometric data from a Fitbit fitness tracking device worn by the project PI/poet to its custom-built poetry generator. This generator utilises NLG techniques to output poetic text published in real time, and 24/7, on the
Eververse website.

The form and content of the poetic output is designed to change according to different physical sensations and experiences in the poet’s waking and sleeping life. Following Charles Olson’s injunction that “the line comes...from the breath, from the breathing of the man who writes” (Olson, 2008),
Eververse’s poetic lines decrease in length as the poet’s heart rate increases and breath contracts. Similarly, the response to the randomness of the dream sleep (REM) state is an increased irregularity in the poetic form. Content, too, reflects these variations, as heightened-sentiment vocabulary is produced to reflect the emotional intensification of an increased heart rate, while the dream sleep (REM) state generates
surreal images and vocabulary.

Technical description

The
Eververse application consists of three main modules. The first module interfaces with the Fitbit device and its data through its Application Programming Interface (API). The activity data of the poet wearing the device is then sent, in JSON form to the NLG module referred to as the ‘generator.’ This generator carries out a number of steps in order to generate and return a poetic couplet based on a conceptual model of states based on the activity information contained within the passed JSON data. The number of words and the frequency of the generated couplets correlate with the heart rate of the poet, whereas the textual content of the couplet is generated from the input corpus which is fed to the generator. Thematic links exist between the content of the input poetic corpora and the conceptual topics of physiology (broadly), and sleep and the body (specifically) corpus comprises poems on the topic of the body; all poems are previously published and none is composed by the
Eververse poet. In order to disassemble and reassemble the corpora for publication, they are arranged in a reverse ngram matrix and further shaped into a frequency lookup table by Poesy, a Markov Model-based Natural Language Poetry Generator. The lookup table is used to create verse lines and a python library,
Pronouncing, is deployed to rhyme the verses. In short, our method takes a language model approach similar to (Barbieri, G., et al., 2012) although we do exploit some semantics, specifically alignment of couplets with Fitbit activity states.

The generator is written mainly in the Python programming language using the micro web framework, Flask. It consists of a web interface to display the generated poetry and an administrator interface that is used to define heart rate parameters for different zones and to determine the form and content of the verse that corresponds to these zones. The extensible approach we use to build the poetry generator means the project can easily incorporate additional biometric data types along with their associated corpora in future.
The public user interface created to display the generated poetry relies on a number of Open Source JavaScript libraries. These libraries enable display of generated text (Handlebars.js, Textillate.js) and retrieval of data from the web application’s API and user interface animations (jQuery). The dynamic background images are created in realtime, and utilise the activity data as an input to affect their appearance, representing a visual correlate to the generated poetry.
Multiple versions of this interface were created for deployment on the web, in a live performance environment, and for display in a standalone exhibition setting. Each interface was adapted to take into account the context in which it would be experienced, for example, differences in how, or if, user interaction was required, and addressing the differing requirements for text size, line spacing, and overall page layouts.

Conclusion

The presentation of this paper will report on the technical work completed to develop Eververse, while reflecting on the implications of the project for issues in poetics, authorship, and automated literature generation. In addition, the presentation will describe deployments of the project in web, exhibition, and live contexts, concluding with a brief live demonstration. Accessible supporting materials for the conference presentation will be made available in English, Irish, Italian, and Urdu
.

Bibliography

Barbieri, G., et al. (2012). Markov Constraints for Generating Lyrics with Style.
ECAI’12: 115-20.

Goldsmith, K. (2016).
Wasting Time on the Internet. New York: Harper.

Olson, C. (2008). Projective Verse. In Cook, J. (ed),
Poetry in Theory: An Anthology, 1900-2000. Blackwell, pp. 288–95.

Reiter, E. and Dale, R. (2000).
Building Natural Language Generation Systems. Cambridge: Cambridge University Press.

Tonra, J., et al. (2019) Eververse. https://eververse.nuigalway.ie/