A couple of years ago, I noticed that you can find something floating around the Internet called a travesty generator. This is a piece of software that makes word salad out of whatever text you give it. The software chops up sentences and paragraphs to form new streams of words, without worrying at all about grammar, punctuation, or proper capitalization. Generally the programs accept input of unlimited length to produce randomized output of a specified length. This can result in a high degree of repetition if the input text is shorter than the output text. The result is unreadable nonsense: a travesty. But there is sometimes something helpful in letting that happen to text you care about, because it gives you a feeling of tone, disconnected from more intentional meaning. I'm told the process has something to do with Markov chains but I've never studied those in detail--although I have made note that the same technique here used to destructive effect on coherent text is being used extensively within Google to generate meaningful search results.
In reading about travesty generators, I thought: Travesty is an interesting word. It can mean something specific and technical about a relationship between source material and a derivative work, which is how it is being used in the software description. However, the word can also be used in a less technical sense to give an event the same connotations we might otherwise describe with words like "debacle" or "fiasco." Could I use a travesty generator to produce text which was a travesty in both senses of the word?
Years earlier, the television show NewsRadio had done something exactly like this. Search the Internet on
Jimmy has fancy plans and pants to match and you'll probably find the relevant parts of the episode. In that story, Jimmy James spoke that infamous nonsense phrase in English in front of a Japanese-speaking audience. He was reading from a version of his memoir that had been translated into Japanese and then back into English, in the hope that what he read in English would more closely resemble the Japanese translation of his memoir which he was trying to promote. The result was something that was both a travesty of his original text and a humiliating debacle for him personally. The double-travesty emerged from a good-faith effort to translate the original, presumably coherent text into a very different language. Here, the slow de-rezzing of meaningful content into funny but content-free language is an artifact of bad translation, and by using that to humorous effect, the show made an observation on the reality that idiom can't always be directly translated between languages.
Around the same time I started thinking about travesty generators, I also became aware that Google Translate had an interesting property which emerged if you cycled a single text through several different languages in succession and came back to the original language: the words got a more rigid, excessively formal tone, and in the process lost a lot of their original meaning. This reminded me of something that Henry Rollins had written in Black Coffee Blues about his experiences speaking English with the Dutch:
The Dutch have mastered the deadpan reply. That's the one where they make you feel like an incredible asshole. No matter what you ask, you will be answered as if someone is reading to you from a book on Russian history. The more energy you put into a question or a greeting, the more you will be halted by slow, measured speech which often contains better grammar than you will ever possess. If you make a joke, the Dutchman will retreat ten big steps down the hallway of ultra-infinite cool.
How much of that is because the Dutchmen who were speaking with Henry Rollins wanted to make Rollins feel like an incredible asshole? How much of it is because native speakers of the Dutch language find the structure and vocabulary of English easier to master than its more colloquial idioms? These Google machine translations invite an answer to that question, although (computers not being people) perhaps not the right one.
At the same time as these translation technologies were emerging into Google's Permanent Public Beta cycle, there was something else going on: a group of researchers had become frustrated with what they saw as the low editorial standards in both scientific journals and in magazines of literary criticism. These researchers produced machine-generated text that was designed not to hang together as a coherent argument or science of any kind, but which had carefully polished for a certain structure, vocabulary, and overall tone. They submitted these bogus articles (essentially a form of textual glossolalia) and then hooted and crowed and shat in their hands and flung the poop of recrimination in the direction of the editors who eventually did publish the material.
I thought: Here was a case where researchers used machines to make something that sounded credible without having meaning, because they hoped to make a point about humans who are supposed to curate for meaning in deciding whether or not to accept articles for publication. Was anyone talking about whether or not we could use those same techniques to get machines to actually create new meaning?
Meanwhile, Internet memes were circulating which pointed out how good the human brain is at picking out meaning from confusion. Terrible misspellings and manglings of words can happen without obscuring the meaning of a sentence as long as the first letter or two and the last letter or two of the damaged words remain intact.
I thought: If that happens to us with letters in words, does it also happen with words in sentences or paragraphs?
Translation is hard. Distinguishing well-formed language from well-formed thought is hard. Machines are getting better at producing well-formed language, and by way of apologizing for the parts of that process that are hard, might turn badly-formed English into better-formed Spanish, and might turn that better-formed Spanish back into better-formed but dumber English.
I thought: Why not give that dynamic a push?
So I took some text from the Internet, a 500-or-so word review of a play. The review had annoyed me for focusing on what I felt were the wrong parts of the play. In this context that appealed to me because I was hoping I could apply software tools to the text and get the basic sentiment of the review wrong. I ran the review through a travesty generator to get it up to about 5000 words, and I fed that into Google Translate through a sequence of different languages: some Romance languages, some Germanic languages, some Asian, eventually back into English. I took the result, fed it through a travesty generator, and repeated the process until I had something that was no longer obviously connected to the original text except by a few proper nouns that survived all forms of translation. The result was incoherent, but it was also full of fun little machine translation artifacts. For example, the presence of French in the translation chain took original English words like "street" and "avenue" (from a part of the review taking about where the theatre was located) and brought them into French as place names, but didn't always bring them out as place names. Sometimes the translation was confused by a false cognate between English and French, taking "rue" as an English-language verb that did not require translation. Other times the translation brought rue back from the French as "the names of the streets" or something similar. The stilted, repetitive interpolation of these two translations in the output text gave some of the passages a haunting, regretful tone. It was as if a person fluent in another language were trying, somewhat unsuccessfully, to articulate something complex and ephemeral about how places influence the relationship between artists and their audiences. The original review had just made some remarks about where the play was being staged, how the reviewer and the audience responded to it, and what the play was about. The shifts in tone and topic were emergent properties of the translation cycle, and hadn't been present either in the original review or in the English-language travesty I had generated from it.
I started piecing the text back together line by line. I deleted strings that hadn't passed spellcheck and strings that still looked like obvious travesty generator output. I deleted long duplicate lines and phrases, and then reordered the remaining passages while trying to preserve the overly formal grammar that Google Translate was giving me back in English, with an overall goal in mind: Add very little to the generated text, delete the minimum possible, preserve the order of the machine-generated text to the best extent possible, all while applying a human-intelligence "de-travesty" process to the jumbled text, doing to whole sentences and paragraphs the same things my eyes do to misspelled words and ungrammatical sentences. Finally, I unfucked the punctuation to turn what looked like a jumble of unstructured free verse without line breaks into actual sentences and paragraphs.
The result was largely incoherent, with sentences like:
What is now little room for optimism, and repeatability of a general truth, and the last five years, is this influence: If gold is very predictable, I see not only a number of languages, and back along the corridor. I laugh, but I DO think. I bow before me. Go home, go to the people.
But there were flashes here and there, stuff that surprised me with its apparent emotion. I thought: what happens if I cut more, and rearrange the results freely? Can I create meaning where there wasn't any?
From that original review, and the misuse of other people's software, I got:
We are game-based nonsense.
In the Shakespeared day, I think forward to industry as the throat of the country, minutes of reform and the day to be.
We are a guest in the great moments; laugh, or at least play.
Rue reform and grooming, fair repeatability, the first-minute impact of cartoons.
Be refused the absurd.
Shakespeare is proud to return home.
Stirring, without saying much. Stirring enough that it said a lot to me without any apparent intent on the part of its author, whoever that author could actually be said to be. If I like (and quote) six random bars out of a sixty-four bar rap song, the rapper still wrote those bars and nobody will compliment me for anything except perhaps my taste. I liked the way these lines fit together, which perhaps says something about my taste, but I certainly didn't write the lines themselves. "We are a guest in the great moments" is one of the more profound things I've read all week, but nobody actually wrote it. It just washed up on the shores of a translation buffer doing its best with fundamentally incoherent input. Claiming credit for happy accidents is part of our intellectual property culture, but claiming copyright on derivative works that were used without permission is thinner ice.
Anyway, I sent the whole generated text to a few friends along with the resulting "poem." My friends seemed to enjoy the result and pointed out some words and phrases that they liked themselves:
- Shakespeare's Dirty Vaudeville-Style Spirits (I'm pretty sure a nodeshell with that title would be enthusiastically filled by someone.)
- Street price is low and laughter high.
- Quality is my true poet, if you have five.
- Is this a great ghost hole?
One went on to write that Jeffrey Jones and Tristan Tzara would be proud, and gave me a few paragraphs of his own on the idea at hand, how (as the author of the play being reviewed, as someone who had read the review which quoted his play at points, and as someone who had read both the "unabridged" and the "abridged" results I had generated from them) he was fascinated by the interplay between various textual sources which were still apparent to him but had mutated to produce surprising results.
My reply to him started with the following:
I took your email, ran it through a fifth-order travesty generator, translated it iteratively through every language starting with the letter "P" and then back to English, ran it through an eighth order travesty generator, translated it iteratively through French, Spanish, and German, before bringing it back to English one more time. German portage of Spanish grammar looks pretty close to ours for some reason. Then I re-ordered, paragraphed, and punctuated the results to resemble sensible language, adding and deleting less than 2% of the resulting text. I kept the ordering of the final machine output to the best extent possible. After just over two hours in front of the keyboard, I present to you a manifesto on a new literary movement: New Method Adaptation. Google produces no search results for that quoted string, by the way. I think we may have a winner.
The rest of the email was a long manifesto, ostensibly written by no one. I don't want to post the full manifesto here at this point, because the text's unintentional co-author has not yet completed his own more serious discussion of creative adaptations. I will include some snippets that seemed insightful to me given that neither party in the email exchange had said anything to that effect:
- I think some of the impact with poetry is what happens when a person is not clear.
- You even believe the impossible is interesting.
- People even sense the infinite, in time.
- How can the source of the text affect--no, reduce--the ideological significance of the materials?
- The art of all poetry is living, interestingly.
I particularly like the way that these results don't just capture the effects of machine translation on a source text, but also capture the thinking of the translation site's programmers at a particular point in time. As they attempt to make their product more useful for its intended purpose, they are changing how their product will behave when I use it for this unintended purpose. Good translation should preserve intent and meaning, more than word choice or sentence structure. The surprising results I got in 2010 may not have been possible using the Babel fish technique that was described on this site a decade prior. Will the results of the same sort of process a decade or more from now feel more or less like poetry? More or less like essay? As the translation techniques start to make better correlations between the idioms of different languages, will the results look more like plagiarism of contemporary authors, or less?
Perhaps most importantly, does what ultimately happened here deserve to be called art? I shoveled text into the hopper because I was curious what would come out, which is the act of a scientist or an inquisitive toddler more than it is the act of an artist. Following along behind those circa-2010 autonomous subsystems with the textual equivalent of a wheelie can and a push broom certainly did not feel like the act of an artist. The choices I made during editing could be said to be a form of intent, but I certainly did not have a direction in mind when I started editing, and did my best to let whatever meaning the overall pieces had be shaped by the machine-generated ordering of the results, rather than my own wishes for the outcome. That the finished product contained elements that were moving to me in some way suggests that art might have happened, but if so, who was the artist?
My friend says it was me, which is gracious of him, but I've felt just as moved staring at clouds in the sky, or watching waves lap at lakeshore near sunset. Am I merely a guest in those great moments, or am I their author as well?
I'll close with one final passage from the New Method Adaptation manifesto which seems applicable to this topic:
However, to talk of and analyze the main ideological point, I say that the nature of text is written to send a request. To create through word and poetry because you can feel each level of the enormous hat tells of a clear sense of recognition, and of the break with contact. Although the resources to develop this document clearly reduce my brain capacity, they say it is necessary to write this that way. For me to send the status of this study, what I want is important.
Don't ask me what the enormous hat has to do with anything, though. I'm just the editor.