about

For my project, I was interested in language on a purely phonetic, bodily level. 

One of my obsessions before this class was Christian Bok’s Eunoia— the idea that certain vowel sounds inherently carry certain characteristics, and specifically scaffold certain world languages more than other. 

In his writing, you can really see this with the flowy French phrases in “Chapter E,” and the terser Arabic vocab in “Chapter A”.

Eunoia took Bok 7 years to write “manually.” I wonder if anyone’s taken a stab at writing this with computer procedure.

There’s such a material quality to spoken word—which are just meaty vibrations, with different ways of stopping air. My 1-month bare bones knowledge of linguistics tells me that air stopped with my lips, versus with teeth, versus tongue, are the most low-level differentiators between sounds that are innately “aggressive,” or innately “soft.”

For me, the exciting computational aspect was glomming together intense similarities in record time— creating this hyper-active form of English that almost evolved into something more abstractly shaped, like music.

procedure

See the source code here: GITHUB

Thank you to Allison Parrish for code help, as well as to Ayal’s project which uses pronouncing + spacy. I referenced code for my "rhyme()" function from there.

~~~~~~

I used the pronouncing library to riff off of seed phrases, which were mostly chosen on a phonetic level:

How well would they amplify the effect of the rhyme?

The main source code borrowed code from the Pronouncing notebook. I put it in a separate function; 

Here’s a few examples of this function in action:

Output:

This one ended up being problematic in its rhyming... I was really upset that it kept returning a racial slur..... Which with Allison’s help + word.filters, I got fixed.

Most of the other poems follow this formula too. 

One that stood out as different:

“Rain in spain,” was the only string that has some kind of intentional cultural reference. Originally from “My Fair Lady,” it acts as a kind of shibboleth for lower-class dialects; there’s a whole song about her conforming her vowel dipthongs.

Seemed effective to riff off of something well-known?

performance

(will post pictures of performance once I get them!)

Overally, I was happy with my feedback on the performance. The pacing was good, and took practice to get there.

Got helpful feedback from Allison and others about not rushing through the language; the contrast in pacing allows the sound to become abstract. When meaningful semantic moments push through, there’s a shift in attention to the “word-ness” of these sounds.

Our guest critics offered that it might add interest/meaning to use a culturally familiar “base” phrase, like I’m doing with “rain in spain.” Could be a nice way to push it further. Or create some kind of narrative? If I could add meaning to this beyond sound...

For sure, all this rhymey stuff I’m doing already has its immediate place in pop culture: tongue twisters, hip hop music, to start. Sofia mentioned that this was a very “American English” performance. Could be interesting to explore different subcultures/languages of rhyme.

“If the cause of the machine’s erratic behaviour becomes known, the glitch tips and becomes a simple bug report of a failure, in which it be described under its technological name (which at that point is often a compression artifact). However, if the cause of the glitch remains unknown, the glitch can either be ignored and forgotten, or transformed into an interpretation or reflection on a phenomenon (or the memory thereof) defined by a social or cultural context (conventions, histories, perspectives) and the technology that is malfunctioning.”

The Glitch Moment(um) by Rosa Menkman

I initially wished to combine this final with my Nothing Final... I learnt how to use the Twitter API with python, extract tweets…attempted making new “phony” campaign slogans for environmental activism using the semantic walk function and word vectors… as well as attempting to play with Genesis 8:22 (“As long as the earth remains….”).

Yet I did not encounter any results I was satisfied with and set out to displace data in jpeg images using text following up on one of my assignments.

The source text I have used comes from all the messages I exchanged with a person from 2013 to the present. It took me some time to assess how best to retrieve this data. Initially I attempted using THIS script -  And got a series of versioning errors in the Facebook developer’s API. I realised only afterwards that I could download all of my messages directly from Facebook in the form of json…. and cleaned and parsed out the content like so:

I am interested in the manifestation of digital ‘memory’ within data and its parallelism to ‘organic’ memory. I have attempted to visit this notion by extracting representations of the memory of myself and this other person from the text without being too explicit. I tried a number of things to analyse the text from using word vectors and markov chains, and settled with extracting the noun chunks from the text using SpaCy- it does not only extract nouns but other small snippets of the content of our messages that I found to be rahter poetic.

In the meantime, I had also understood via my previous assignment that in python 3 it is possible to open the bytearray of a jpeg file, in addition to encoding and decoding strings into bytes, hex or ascii. And that it was possible to replace the data of a jpeg file with it - thus corrupting or “glitching” the image.

Through my quest to make a statement about my remembrance of this person, and to somehow interpret the phenomenon of our relationship over the years, I wished to subtract things I had said from things they had said through the manipulation of our data- hence the name or the project being “The Subtraction of Me From You.”

In order to “subtract” myself from B, I extracted the noun chunks, transformed them into strings of characters - translated each one to the corresponding ASCII character. Therefore acquiring two different lists of ASCII characters on which I performed the following mathematical operation:

I then proceed to transform this character list into a bytearray that is written over the original data of the JPEG file.

For the purpose of performing this program, I wish to subtract words from words, and the result of this “subtraction” to be the real-time display of this glitch. I want to “subtract” myself so much from this person that eventually the image, which will be an image intrinsically valuable to the memory of our relationship, will be corrupted entirely and can no longer be opened.

I have proceeded to transfer my notebook to a py file that I can run in my terminal. I used the input function in python so that I can decide the  amount of noun-chunks to devise my poem that then wait for a prompt to use the data to overwrite the jpeg. In order for the image to be seen in real-time, Leòn suggested to input the image into an html file that gets constantly refreshed in the browser as I manipulate it.

I wrapped most of the proceedings of my script into functions in order for this to happen & included a counter that makes sure the data being replaced is at the beginning of the bytearray of the JPEG, where the header of the data belongs.... in this way I can be sure that the image is corrupted at a certain point during the performance...

Had I had further time time to work on this final, I would have ideally liked to make it so that the eventual breakage of the image was caused by another operation I would perform using the text to ascii conversions. I have tried doing this, but it would eventually take a lot of time for the file to eventually break- thus I have negotiated for the sake of the performance.

This week, I played with word vectors and sentence similarity. I don’t know if I built anything that feels “finished” yet, but it’s been fun. 

Here are two pieces in progress:

~~~

1. Facebook Senate Hearing Blurbs

I watched this happening on and off this week, and still have yet to digest it more thoughtfully.

I thought it’d be interesting to play sentence matching, and how folks felt Zuckerberg robotically shirked responsibility in every answer.

I sourced the full transcript from the Washington Post. I used spacy and word vectors to find similar word tokens:

Looking for interjections/defenses starting with “Senator...”

Looking for laughter, or something like it?

I combined a few of these to create an abridged reading of the hearing. This code is looking for, “But is it safe?” and “Laughter,” and “That’s not what I asked.”

~~~

2. Who do Aliens look like? 👽

I’ve been interested in this idea of “otherness” in science fiction.

There is some great writing out there which points out how characters of color are described so weirdly in literature, or what it means to be Alien American from a vaguely “oriental” planet; how galaxies need more humans of color.

My source texts are from Wikipedia’s listings of extraterrestrials in popular culture. I used Star Wars, Stargate, and DoctorWho, because they’re probably the most popular of the popular.

You could definitely argue that the aliens might have been very fair in complexion, or brightly colored, but got averaged into a browner hue. It’s definitely fair to question the “accuracy” of this color-averaging!

Here are some descriptions of alien people using these new colors. 

Supposed to be satirizing, and hopefully does not feel offensive!! <3 I wonder what I can do to make that clear.

My findings have been interesting. I am pulled in different directions by intentions, meeting assignment criteria and letting python take me where it wants to take me as opposed to tugging at its hand with impatience and desire of control.

I thought generating markov models with a terribly formatted corpus of text of messages written between myself and a friend between 2013 and 2018 would be interesting - - due to distance between us, our relationship went through many stages through written words and I wanted to compare what we wrote to each other once and the messages we have shared recently. When i simply tried to feed this into the gen_from_model function I would get some interesting results.... We went from sounding like this:

to sounding like this (the last three “I am” s in this I found particularly interesting.

I was simply trying to get a feel of what the chaining could do, but I feel like the original text should be formatted better in order to train it better. Additionally, if I went over 2 n-grams, it would lose this nice subtlety of re-contextualization.

In any case, I was not satisfied with this; I kept having the urge to bring together two inputs for the markov generation and stubborn to continue using this text. I wasn’t sure whether to apply spacy before or following the generation of markov sequencing and played around with that, with again results that did not provide any syntactic, phonetic, formulaic, orthographic delights. That’s when I thought it would be interesting to bring in something so far away from what I familiar with... When I was in highschool, I did not code but I found out through the internet how to data-mosh by messing with the data of a jpeg file via a texteditor. Sublime Text opens jpeg files in what I believe to be hexadecimal (or binary - to be honest I have been looking at encodings notes online for a very long time and I am still somewhat confused)...But I extracted this data into a text file and ended up not knowing what to do with it either...Until I found this website called devdungeon.com and found a mapping function that I used to encode text into binary data in my notebook. I wanted to take the data from the image of this friend and I, as well as the corpus of text that I was using previously and find some kind of relationship between them to then “glitch” or “corrupt” the image as a means to symbolically represent the distance that has grown between us via this platform. I used spacy and regular expressions to extract the sentences in the corpus that had specific words in them... For the sake of trying this out, I used the strings and tried converting them to different encodings (playing around with ascii, hex, utf-8, bin etc) and copy & pasted them into my original jpeg file open in sublime. Converting them into binary values and then replacing some of the values in my jpeg file worked! It would glitch the image, that I obviously did not save and then corrupted too much. I wanted to corrupt the file differently in accordance to the specific extractions of text I had made using regex and spacy. I thought it would be interesting to also see if I could translate text into binary data, then feed that to the gen_from_model markov function and then try to decode it to human text again. I did not figure that out but that’s probably because data does not work in the way that I think human language does. It is something I would like to explore further for the final I think. T Anyways at this point I really have to figure out whether what I am doing is entirely possible as I decided to take these values and re-inject them into the original image file by randomly replacing specific elements of the list ( a jpeg file has some specific protocols that it needs to remain a jpeg file) and I corrupted my files too much that they would not open anymore...I tried d On the other hand, the patterns being made by this thing that I’ve built were extremely visually pleasing to me:

Here is the gist of the most complete of my notebooks:

This week has been a little crazy, so apologies that this week’s assignment feels a little all-over-the-place! Instead of making one fully interesting text, I played around with a few different source texts through one Markov generator.

This week, I spent most of my time reviewing Spacey, playing with neural nets (I haven’t gotten it to work yet), and finally, feeding .txt files through Markov models to generate some weirdness. There was so much precedent to understand first.

Out of my tests, I think #1 is successful; I think #2 could be interesting but isn’t, and #3 isn’t interesting. More about that at the end.

Here are the texts I played with:

1. Apple’s TERMS & CONDITIONS.

Taken from here. Why? 

Because I was interested in the convolution of legal language which could very well be generated. It’s barely parsed by the average human being. It pulls on a library of highly constrained legal lexicon that can just be reiterated ad nauseum and still sound the same to humans. And, it could say literally anything and most people would agree to this texture of text, oblivious of meaning. 

(I always have a fleeting thought that I might be donating my kidneys to Apple by checking a box; I don’t actually know what’s going on.)

Here’s a snippet of the original:

Fnal output, with a parameter of 2:

It just becomes this texture of legal grammars and jargon; it doesn’t feel all that different from the original, even though it is:

Terms and conditions, not this agreement, and any billing disputes related to your email address or a book, apple distribution international and you hereby release apple from any and all liability to you, and therefore apple is a subscription music service. your apple id. please contact apple at the same home country. not all content, including in-app purchases, subscriptions, and some previously acquired apps, are eligible for purchase sharing. e.

But also-- it’s funny to me. “And therefore apple is a subscription music service.” Another:

You will be uninterrupted or error-free, and you are not a party to the recipient. no apple employee or agent has the authority to vary this agreement. if (a) you are not located in the following paragraph, this agreement or as a “transaction.” 

Uuhhhh, yes I agree. Now let me use your products!

In seriousness, I think this is one of the more interesting experiments. Legal language, because it is so formulaic and often modularized, is easily predictable. At least with Markov.

This could easily be extrapolated to other legal documents and have a similar feel.

My neural net file wasn’t working, but I think it would have been even funnier with neural net processing-- nonsensical occurrences at the word level would have lifted it into some intoxicated, surreal state.

~~~~~~

2. Bob Dylan lyrics. 

Why? In my high school days I practically memorized some of his 6-12 minute rambling epics; they feel like they could be reiterated ad nauseum too. They’re rich with surreal, sometimes cryptic imagery; they can be formulaic and repeat forever. He’s a very, very generative writer. 

The original lyrics to Visions of Johanna can be found here.

The gif below is kind of lossy: basically, it’s demonstrating how the length of the tuple in the model influences preservation of the original verse:

How are these affecting the output??

___ Parameter of 1:

“ain’t it lights flicker from the d-train but we sit here stranded, though we’re all muttering small talk at much, can hear the jelly-faced women doing our best to all me up past the all-night girls they make it....”

 “ain’t it all sneeze hear the mirror but like after a parasite and say a lotta gall to me, he sure got a lot of prayer for him or them that’s really insane louise and johanna now have steps to play blindman’s bluff...”

These just go on, and on, and on. 

Having a shorter chain element can get you into recursions a lot faster. The grammatical & semantic consistency is more compromised, so I would say the sentences are less interesting-- a kind of word soup is created.

“can hear the jelly-faced women doing our best to all me up“ (?????)

In my opinion, this really isn’t too interesting.

___ Parameter of 2:

“aint it just like the mirror but she just makes it all seem so cruel. ”

This is short and sweet. I think this could very well be a poem; you’d leave people wondering with the comparison (what is cruel about a mirror?) about what’s going on in the unsaid.

The rest of the paragraph feels very close to the original, with more random recursion mixed in.

I think this is telling; Dylan recycles choruses and rhyme structures often, but doesn’t repeat unique words all the much. Which is also NOT as interesting for predictability.

___ Parameter of 4:

“ain’t it just like the night to play tricks when you’re tryin’ to be so quiet?”

This is the original, and so it the rest of the paragraph. We’ve reached a threshold of non-change.

Perhaps this would be more interesting on a neural net. I think made-up words and surrealism could be a fun add to an already vivid set of lyrics.

~~~~~~

3. Ursula LeGuin’s speech at Bryn Mawr. 

Why? I didn’t have interesting rationale, other than 1) the speech is about the power of language & grammars  2) I really love everything about this body of text.

Read the full thing here.

It has enough repeats to branch, fold in on itself, do interesting stuff.

I think any predictive procedure to the essay muddles the expertly crafted meaning. The narrative gets destroyed, and I don’t know what might be interesting about a statistical analysis of this text.

Recursion COULD add something to a performed speech, maybe. I don’t think it adds much here.

“the language of the seventies and eighties - I celebrate here and now the women for who two centuries have worked for our freedom, the unteachers, the unmasters, the unconquerors, the unwarriers, women for who two centuries have worked for our freedom, the unteachers, the unmasters, the unconquerors, the unwarriers, women for who two centuries have worked for our freedom, the unteachers, the unmasters, the unconquerors, the unwarriers, women who have at risk and at high cost offered their experience as truth.”

I also don’t think a neural net would make it more interesting.

~~~~~~~

reflection

I think these random experiments are telling (for me) how predictive text adds or detracts... at least in these Markov scenarios.

The Apple one feels successful because it thinks about infinite text as sheer mass; there’s a uniformity; meaning is rendered unimportant. Legal text also modularizes nicely; it’s not a particularly sensory or humane language. The drone of a computer making all this writing is also a weird gesture.

For this assignment I was initially interested in creating a poetic form that involved abstraction of syntax as well as repetition. I was excited to use tracery and eventually found that it did not fit the needs of what my experiments amounted to. I also decided to implement some data that I consider to be somewhat sensitive and weighted in nature. I have been doing some personal research on human trafficking in Europe; in my own country in the outskirts of cities like Rome and Milan there are locations that have become the workplace and often home of many prostitutes typically from Nigeria and Eastern Europe. Many of these Nigerian women are promised work and papers and brought to Italy with an immanent road to prostitution waiting for them. They are placed under the   direction and care of a “maman” (usually another Nigerian woman with the same past who has paid off her “debt” to her previous mistress and pimps and become one herself), who administers their sex work. Without going too deep into this, essentially this figure of the maman is inspired by a water deity venerated in West Africa (and other parts). This has been translated by the Nigerian mafia as inspiration for a process in which the girls are introduced to their circumstance. They must undergo an initiation in the form of a voodoo ritual to their maman, in which three promises must be made involving devotion to their mistress, an oath of silence (about who brought them to the country) and the payment of their debt (for their travels).

Through this research I also came upon a website called Doxyspotting.com: “This website gathers international Street View sightings of sexworkers, brothels, pick-up spots and red light districts. We are the biggest archive of prostitutes and prostitution-related Street View sightings on services such as Google Street View, Bing Street Side, etc.” (http://doxyspotting.com/about/)

I set out to take the usernames of the “spotters” on this website which are placed in ranking on the website in accordance to how many “spots” they have contributed to the website and use them in conjuction with the number of times they have tagged sex workers. So, I ended up spending a large amount of time trying to clean up the data set I had in the form of a list in my notebook--understanding the right comprehensions to use as well as having tiny hurdles involving syntax... The data set looked like this:

I had a few creative blocks about how to use it-- emotionally this was difficult and technically I felt at this point that my knowledge of python wouldn’t allow me to create something worthy of this subject matter- but I persisted with the idea that, inspired (if that’s the right word) by the tainted theological process I mentioned previously, I wanted to create a poetic form that could be applied to this data set to create curses or omens towards the users in accordance to how many times they have “spotted” .

I wanted this form to be aggressive through syntax and through repetition. I looked through corpora for some data sets I could use; I didn’t know if using insulting words would be ideal or something I could play with---> I decided to use a list of dictionaries with different codes and diagnoses. I transferred the values of the keys in these dictionaries into a list called disease_data. I wanted to use the capital of each username to create a text that would only use that letter for the beginning of each word in that sentence, but the disease_data I found came in form of sentences and took more time to extract therefore had to compromise and initially just used that. I used for loops and comprehensions as well as f strings in order to extract a random username and the number associated to that “spotter” and this list of functions to extract a diagnosis beginning with the same letter...

Eventually printing out the “spotter”’s username with the diagnosis printed the number of times they had spotted--> some of my initial results look like this:

I was not satisfied even though this took me some time... I would have liked to create something more complex...Therefore tried using tracery to add further abstraction ( I wanted to implement an adverbs json file and apply the same first letter rule to that list but haven’t had the time therefore simply added a few other rules into the dictionary )

The form in which I printed this felt too ordered; I tried reading this out loud and found that having a line break after each diagnosis with a capital gave too much of an ordered nature to the way in which this curse was being summoned...therefore I attempted to apply further line breaks and a repetition of the username and this is an example of the results:

There are problems with the way that my data is being read such as the diagnoses not being randomised for each letter, and my username-number of spots data set not being split very well as each element in the list does not have only two elements... I feel that the only successful part of this assignment has been the repetition in the form...I am not sure if this is the right way to approach this data but it is an attempt that has driven me to want to explore this idea further and solidify my intentions.

For this week’s rwet class, we were tasked to think about forms of poems— the underlying structures of them— and to make our own.

For mine, I was inspired by tongue twisters. These short poems are extra interesting to me because they can be read pretty smoothly by a computer; a human is made of meat still and can’t make movements of tongue and teeth fast enough. 

🌶〰️🌶〰️🌶〰️🌶

process...

To proceed, I picked the structure of “Peter Piper” and put it into Python. I selected some spicy source material from hot_peppers.json, taken from Corpora. 

At first, I wanted to write a script for alliterations across a body of different .jsons, but felt a bit intimidated starting there. Also, I wasn’t sure how to control for words which started with the same letter, but didn’t actually alliterate well. (“Jesse James jostled a jolt of jabbering jalapenos”? Nope.)

In my initial version, I manually generated alliterating names, verbs, nouns for each peppers that I felt would “sound tongue-twisty” in regards to each pepper. Then, I fed it through a tracery structure with base_english grammars. 

“Dane Anders probed a bag of piddled medusa A strain of serrated tabasco Tobias Fran did anchor If Kayla Casava pacified a franchise of pickled cayenne, How many medical serrano did Fresca Banister frequent?”

In summary, this sounds like an algebra problem from an alternate universe, where internal consistency goes out the window.

🌶〰️🌶〰️🌶〰️🌶

perfect, progress! 

So granted, I did want the consonants to alliterate nicely-- that’s often the whole point of a tongue twister. 

I used list comprehension to create new lists of only words from the first letter of a specific pepper. Like so:

Now what? Next, I fed it in through my existing tracery structure. And voila:

Still images:

~~~~

~~~~

~~~~

~~~~

The structure borrows from the original Peter Piper, made computational; unlike Peter Piper, it re-randomized each line. I made my tongue-twister longer as well, in so that the reader could enjoy some of the variation.