Convincing (?) Art Historical Text Generation

It’s a commonly held prejudice that analyses of art are arbitrary; you can look at a piece of art, repeat any stock expression, and your audience, no matter how educated, will buy it. So when prompted to address contemporary art in some way, I decided to test that theory by submitting ten pages of randomly generated art historical text. Producing text with a Markov model raises questions about the role of contemporary art scholarship. On the most naïve level, it is a humorous exercise in seeing the actual words that art historians and critics produce. On a deeper level, however, the amount it convinces us of its veracity or potential veracity provokes two kinds of thoughts. The first, critical in nature, prompts inquiry into the extent that art criticism is derivative. If this text, which is by construction the shuffling and reprinting of previous texts convinces us to some degree, then we must wonder how “original” art historical knowledge production differs in scope. (Though one likely had this thought before reading any algorithmic output.)

The second thought is reflexive in nature: the position with which we naively approach such text is itself a meaningful metric for how we conceive of writing on contemporary art, and perhaps contemporary art itself. Though obviously silly, the model itself contains a vast amount of information, available immediately with the correct query. As such, it is not inherently dismissible; it knows more about contemporary writing on contemporary art than (arguably) any person, though it may produce nonsense. The position it provokes in its readers (scientific interest, disdain, amusement) is immediately communicative of their baseline thought on contemporary art criticism, and its underlying relationship to the art itself.

Contemporary Art

For example, two chairs in a structure of racial inequality. Is there one? To withdraw from the East River to the general culture. But, if so, you could argue that the internet is based on solidarity, on the refusal to be privatized is our bodies. Yet if police responses to Occupy encampments around the United States; it is arguably the most complicated form is metaphysical guilt: for Jaspers, guilt in the Saturday Night Live skit, The Nerds, though. The look is borderline nerdy. I remembered pantywaist, a word that comes after Art. He has spoken of the exhibition Air de Paris, Paris; Pilar Corrias, London; and Esther Schipper Gallery, Berlin, Germany. On 23 October, his solo show at 47 Canal; New York. No less than a quarter of the participating artists are likely to see. After receiving a Golden Bear at the time, were really talking about Brazilian identity. Cristina Freire: In the case of the philosophy of Deleuze, but clearly favors differentiation over repetition. He offers a sketch or a norm implicit in a book. The carbon chains, O-H3, C-H3, C-OH, gradually penetrate my epidermis and travel through the deep conservatism in neo-formalist work being created by artists much younger than her ever questionable inclusions and exclusions arise, oddball juxtapositions distract.


Contemporary art text generation is based on a statistical model called a Markov Chain, a simple but powerful system used to model structures as diverse as the google search algorithm, grammar, the stock market, board games such as monopoly, weather, and insurance pricing. Markov models are based on three basic assumptions:

  1. There are finitely many states the system can occupy.
  2. There exists a probability distribution on the system that gives the conditional probability of moving from one state to another.
  3. At each time step, the future is conditionally independent of the past.

The final assumption is what makes a Markov model meaningfully different from other ways of modeling multi-state systems. Let’s assume for the sake of didactic clarity that we want to understand how two games, blackjack and snakes and ladders, work. I can do better at blackjack if I can remember the previous cards which have been dealt. Since the dealer goes through the deck incrementally, as he or she puts down cards, the probabilities of winning at each hand change. In snakes and ladders, at each turn, my position on the board and my dice roll completely determine the probability of any outcome. Since previous dice rolls do not affect the probability of future dice rolls (i.e. they are statistically independent), and the only thing previously determined is my position on the board, each turn in snakes and ladders is conditionally independent of previous turns. Snakes and ladders can be Markov modelled; blackjack cannot.

For text generation, Markov models use the assumption that for each word (or fixed-length phrase, confusingly also sometimes called a “word”), there is a probability distribution on all the words that can follow it, which are further assumed to be finite. For example, if I say the word, “MoMA,” I might follow it with, “exhibition,” “curator,” “is,” “galleries,” etc., and each distinct word has a fixed probability of occurring. (The probability of a nonsense phrase like, “MoMA steamroller” will have probability at or close to zero of occurring.) Each pair of words is called a prefix-suffix pair. To generate text, then, we can pick any prefix we want, then randomly choose one of the suffixes that follows it, weighted with its probability, and choose one of the suffixes that follows that word, and so on. At each word, the words that follow the prefix depend only on the prefix – the future is conditionally independent of the past.

Implementation Details

The text generator was coded from scratch based on text scraped from online sources, also from scratch. I programmed the crawler and scraper in python using urllib to make http connections and BeautifulSoup, a screen-scraping library based in XML that translates web pages into trees. (I originally used these packages for a CS problem set). Based on the source code of the website, my code algorithmically generates the archival page urls which contain the urls for individual articles, then visits each article url and pulls the text, writing it to a text file. The model is built by analyzing the individual source text files.My program read the entirety of Jerry Saltz’s critical work for NYMag (about 200 articles), the entirety of e-flux (about 850 articles) and the 100 most recent issues of Frieze (about 6000 articles).


In order to generate text, we first need to build the probability distribution; the data used is called the corpus. The larger and more diverse the corpus, the better the text. For the sake of grammatical clarity, I fixed the prefix and suffix lengths to two words. This produced 275668 unique prefixes, each mapping to at least one suffix. Because the program needs a seed word to begin generation, I also compiled a list of 325 unique prefixes which begin articles. (Random example: “Separations are”.)  To control length and give a stronger impression of coherency, I generated text by paragraph, ending with the last period before a newline break in the original text. Though longer paragraphs give a higher probability of having a diverse range of topics, short paragraphs have surprising diversity. This paragraph was produced with a base length of 10 words:

Why do we think it’s from smoking or the pollution, because it’s too much energy and exist only in a contemporary living room, she did spice it with his kisses. Last autumn, when I was about finding ways to buy it.

The paragraph began with “Why do”, the 99th item in the list of starting words. That prefix has five possible suffixes: “I have”, “we reject”, “a lot”, “America and”, and “we think.” “we think,” the algorithm’s choice of suffix, maps to 12 suffixes, but “it’s from” only produces “smoking or.” Generation randomly oscillates between prefixes with widely varied suffixes and those with very limited options. Every paragraph in “Contemporary Art” was produced in this way, with a base length of 75 words.

To see the importance of a large corpus, we can compare this text to a Jerry Saltz bot. Generation with a base length of 20 words produced, “Jaccuse museums of bullshit! Of bogusly turning themselves into smash-hit consumer circuses, box-office sensations of voyeurism and hipster showbiz. This year, the Dependent debuted in a car accident in 1977, shortly before her 41st birthday. Morton combined conceptualism, postminimalism, magic, Americana, and kitsch with a feminist twist and a comedians wit.” Though equally coherent, this text is inherently restricted to Saltz’s syntax; its ability to produce diverse content is limited. (It should be noted, however, that each of the corpuses – Saltz, Frieze, and e-flux – are large enough to produce coherent text on their own.) As with all data projects, the quality of the text is limited by the quality of the data.

So, did it work?

Somewhat. Markov models, though straightforward to implement, generally don’t give the most life-like text as sentences have a tendency to meander and bare little relation to each other. In the future, I want to experiment with methods for connecting sentences to each other, perhaps by subdividing the corpus into content areas. It seems like NLP/NLG can become arbitrarily challenging, so it would be interesting to see how deep one has to go in exchange for a payoff in readability.

I submitted two documents: “Contemporary Art,” a ten-page randomly generated paper; and “Methodology.” According to my professor, “methodology sounded boring,” so he immediately read the paper and felt it was comparable to what a “bad art student” would “try to pass off as art history” but not really recognizable as an art historian’s paper. Thanks Matthew Jesse Jackson for being a supportive guinea pig and signing off on the project in lieu of actual, you know, art history.


Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

%d bloggers like this: