Poems generated by finding midpoints between how words sound and are spelled

Allison Parrish is one of my favorite poets; she writes code that transforms language into fascinating and evocative literary forms.

Recently she wrote software that generates "compass poems". She'll plot words in the NSEW points on a compass, and her software interpolates new words in the in-between spaces, by finding midpoints between how the original words are spelled and how they sound.

BOMB magazine published a few of them here (that image above is from their site), and Parrish has a whole chapbook of them here, published as part of the sync series.

One delightful one is the Beatles arranged in the compass points, from BOMB

Or, from the chapbook, the big Internet giants …

Or the cardinal virtues of traditional Christian theology …

Parrish's work reminds me of the concrete poetry of bp nichol, a Canadian artist who — from the 60s to the 80s — similarly dove the phonemic muck of language, remixing words into trippy, often-illuminating new combos. Like many concrete poets of the time, he played around a lot with typographic layout, using everything from the electric typewriters to, in the last years of his too-short life, software. It's particularly sad to me that he died just before the web and open-source natural-language-processing tools emerged; I imagine he'd have found those to be deliriously fun tools to play with, in precisely the way Parrish uses them.

At any rate, go read that whole chapbook of Parrish's — it's just terrific. I'm gonna print it up; I want to look at 'em entombed in atoms, holding them in my hands like an animal.

If you want to know more about how her code works, she describes it at the end of the chapbook …

I trained a machine learning model with two parts: a "speller," which spells words based on how they sound, and a "sounder-out," which sounds out words based on how they're spelled. In the process of sounding out a word, the "sounder-out" produces a fixed-length numerical vector, known as a "hidden state," which is essentially a condensed representation of a word's phonetics. The "speller" can then use the phonetic information contained in this hidden state to produce a plausible spelling of the word. The hidden state, like any other numerical vector, can be modified: translated, multiplied, blurred, averaged.

Each of the poems collected here results from a computer program I wrote that performs the following steps: (1) use the "sounderout" to find the hidden state for four words, drawn from a handauthored list (these are the words on the "points" of each poem); (2) find the vector halfway between the hidden state vectors for each pair of "point" words, and predict a plausible spelling for these halfway vectors with the "speller"; (3) find the vector of the midpoint of all eight vectors produced in (1) and (2), and likewise predict a plausible spelling for this vector.

The words from steps (1) and (2) are then programmatically arranged in the form of a compass rose. The word resulting from step (3) is placed in the middle.