Individual Work
Big Data Poetry

Jhave is a digital poet based in Montreal, currently teaching at Hong Kong’s School of Creative Media. Indexed in his grid-like artist statement, he describes his work as “play with language, images, video, theories, dreams, disparity, code, wistfulness, tentacles, food, integrity, time, paths, love and intransigient synchronity.” His practice investigates ontological implications of digital poetry, a problem on which he has published a book, Aesthetic Animism (MIT Press, 2016).

Aesthetically, much of his recent work is indistinguishable from lines of code. For his Big-Data Poetry project this is more than incidental; a deep learning language generator called Pytorch automatically generates the pieces in this body of work. Developed by Facebook, Twitter, Salesforce and a number of other tech giants, Pytorch can be used to build algorithms with minimal memory requirements and self-developing libraries that learn from each output. To build the Big-Data Poetry project, Jhave worked with Pytorch to write his own language generation software in GitHub. With 639,813 lines of poetry from five archival poetry websites as a word library, he wrote code that automatically generates a poem every 20 seconds.

The cadence of this poetry—which learns from the database it randomizes—is eerily humorous, intelligent, and at moments, somber. At the same time, the caesura often feels random. The limits of the machine are evident. Without empathy, and an indifference to context or continuity, the lines of poetry are obviously off. They become more coherent the longer the algorithm runs as each iteration of the software seeks to advance the “quality” of the poem by running analysis on the code’s past output.

What is produced is an assemblage of language given meaning only when the human mind brings sensitivity and curiosity to the software’s words. In Roland Barthes’ fragmented glossary of an autobiography, he muses on the possibility of words which are not read, or written, but received, “I can neither read nor write what you produce, but I receive it, like a fire, like a drug, like an enigmatic disorganization” (Barthes 118).

Adding a sonic layer to the generated text, Jhave chooses words to read aloud with the timbre of a poet accustomed to listening very carefully to the way people speak. Improvising a path through Pytorch’s text, the poet parses the machine’s language, functionalizing its labor and humanizing its output. As Jhave speaks back to the overwhelming indifference and capacity of Pytorch’s text, he runs google transcribe in a side window. His vocalizations of the generated words are quickly worked back into the logic of the computer, returned to digital text by google’s clumsy transcription.

The poetry winks at the overwhelming productive capacity of big data. It openly acknowledges the effortlessness with which machines render insignificant and miniscule the exertions of poets and artists. While algorithms like Pytorch almost mock the human desire to make meaning of stimulus, Jhave shows that it is possible to take words and sounds back from big data. Accepting the limitations of the model, an optimistic Jhave wrote that “in the short-term, this is the future of writing: a computational assistant for an engaged imagination intent on exploring the topological feature-space of potential phrases.”

Author statement: 
BDP (Big-Data Poetry) currently investigates machine learning and neural networks as tools for literary creation. The work is ongoing and open-ended. Previously in 2011-2014, BDP used a combination of data visualization, language analytics, classification algorithms, entity recognition and part-of-speech replacement techniques to 3 corpuses : 10,557 poems from the Poetry Foundation, 57,000+ hip-hop rap songs from, and over 7,000 pop lyrics. Based on these templates, a Python script generates thousands of poems per hour. Sometimes Jhave reads along with this writing machine, verbally stitching and improvising spoken poems.