Authoring text under control: from automatic writing to autocomplete

Code Societies 2018, Days 4 & 5, with Allison Parrish

written by

Session 1: What is text really though?

For three sessions of Code Societes, we are indecently fortunate to have the teaching presence of poet and programmer Allison Parrish. A member of the full-time faculty at NYU’s Interactive Telecommunications Program (ITP), the focal point of Allison’s work is the intersection of the language behaviors and computation.

The first session begins by Allison detailing her own background and some of her own work. Although in disagreement with Twitter’s larger ethical intent, she explains the benefit of this social network in particular as a platform for experimental language aided by computational process. One of the advantages is Twitter’s open and rather encouraging attitude to application developers (including bot makers). One project in which Allison made use of Twitter’s generous APIs is “The Ephemerides”. Every four hours, this bot selects a random raw image from NASA’s OPUS database and combines the visual with computer generated text fetched from Project Gutenburg. The script behind this bot parses texts for grammatical sets and swaps the next section with a random selection of the same set from another book, while random line breaks make it appear like what we are programmed to recognise as poetry.

Allison asks the question, we usually think of computer processes as being mechanical, how can we use them to be lyrical?

Authoring Text Under Control: with Allison Parrish

We consider the properties of digital text:

  • Its speed and scale as opposed to physical media
  • Digital text allows us different units of analysis (bytes, Unicode, files, n-grams, vector) as opposed to just characters and words.
  • It wears heterogenous authorship on its sleeve — digital text is always enclosed within other bodies of text. For example, an essay on a webpage is surrounded by ads, and the interface of the browser itself
  • It’s linear — one unit follows the next — and therefore iterable
  • It’s necessarily a formalisation of text — which forces the question of what to include in/exclude from that formalisation, e.g. when digitally transcribing a handwritten document, does one include crossed out words? If so then how are these denoted?

Allison defines for us two definitions of automatic

  1. What the body does outside of what the mind is asking it to do, e.g. heartbeat, breathing.
  2. Corresponding to a programme, e.g. a dishwasher that performs on its own.

The relevance of these definitions is explained to us in the context of writing. When writing, one typically has an intentional thought which is translated to hand movement which applies marks to paper. In automatic writing the intentional thought is replaced by a subconcious or spiritual process. We look at uses of this automatic writing, such as in spirit mediums, dissociation, subconscious/unconscious, the Ideomotor Effect, etc.

Allison walks us through some examples of our second definition of “automatic” in procedurally generated writing. These include Sea and Spar Between by Nick Montfort and Stephanie Strickhand, To Make a Dadaist Poem by Tristan Tzara (1920), and Call Me Ishmael by Jackson Mac Low. Automatic writing allows us to envision writing that resists control, in turn resisting this A Thousand Plateaus quote “language is not made to be believed but to be obeyed, and to compel obedience.”

Quotes portraying the restrictions of language

Writing is a reflexive process — leading to the big question: how in control are we of our intentional thought?

Session 2: Predictive text and text generation

Our second session kicks off with reviewing some participants’ responses to the assigned exercises from last class. First up is Tanja’s video transcription of a young family member’s birthday celebrations in which a party-goer ended up with a toy airplane tangled in their hair. Annaka transcribed an entirely different handwritten artifact on the Smithsonian’s Digital Volunteer Transcription Centre. Maria presents a transcription of a social encounter she recorded by accident while shooting a street corner for an hour in LA, complete with screenshots of the instagram conversation that followed.

left: Tanya’s dynamic transcription of a home video, right: Annaka’s transcription of a handwritten scientific document

Extending our perception of Twitter as a generative platform, we look to @tiny_star_field , an account which, every three hours, tweets astronomical Unicode characters accompanied by randomised spaces to create, hence the name, a tiny star field. This in itself appears political since the concept reclaims the space of a tweet, proving that Twitter can be used for more than just commercial purposes. Allison further demonstrated this approach with her bot @everyword and the now deactivated @fuckeveryword. These two paralleling accounts raised the question of authorship in a digital age where bots can and will be remixed and republished. How can we navigate others capitalizing on our conceptual effort with a new creation of altered source code?

left: an example tweet from tiny star field, right: tweets from everyword

This rest of this session prioritized understanding predictive text and text generation in a coded context through workshopping the concepts discussed in the previous class. Using Python run through Jupyter Notebook, we explored frequency significance, n-grams, Markov models/chains, and Recurrent Neural Networks. These provided us with the tools necessary to generate our own texts by combining articles fetched from Project Gutenberg.

Actualizing our understanding of n-grams
Twilight vs. Genesis: Combining models generated with Markovify