The Art of Commerce sits on the corner of literature & the marketplace, asking the age old question: Who's got the right of way?
We talk with writers, editors & entrepreneurs about, really, anything. All conversations are 'manuscript-first', meaning they were typed as you see them. Small edits have been made for structure.
Episode V: "Oh, I should definitely explain why I don't care about this question"
In the fifth installment, I speak with Nick Montfort, an electronic poet and associate professor of digital media at Massachusetts Institute of Technology. Topics include the Trope Tank, digital translation, poetry generation, creativity, Samuel Beckett, a Turing Test for poetry and judging generated poetry.
(Andrew Lipstein) I’m here with Nick Montfort, associate professor of digital media at Massachusetts Institute of Technology, though that concise title belies the variety and depth of Nick's work. First things first: we’re halfway across the country from each other. To bridge the digital divide, would you kindly describe your current environs?
(Nick Montfort) I'm in my apartment in New York, where I'm living for 2015, with a glimpse of the East River out the window.
You're spending this year at the New School, is that right?
I'm teaching one class at the New School this semester and am a visiting researcher there. I have a number of poetry, digital media, and academic projects I'm working on.
You’re the director of the Trope Tank, which develops “new poetic practices and new understandings of digital media by focusing on the material, formal, and historical aspects of computation and language.” Can you elaborate into what, exactly, that means?
Sure, I'll be glad to. The Trope Tank is a lab with material resources (old computers, video game systems, physical media) and also supports activities involved with current computing. So, we do research, teaching, and creative projects in collaboration, sometimes working in the space itself and meeting there and sometimes working remotely.
For instance we have done phase 1 of the Renderings project, a literary translation project where we translated and in many cases ported computational poetry and literature from other languages. We'll continue that project, done mostly in the physical lab, with a wider group of collaborators from around the world.
You translate digitally?
Let me explain what's involved.
We have a work of digital literature, usually a computational one, such as a poetry generator. For instance, "Sample Automatic Poem" in Spanish by Félix Remirez.
We use some of the same techniques that literary translators use to translate a poem, but work with this sort of digital text to produce a version in English.
http://curamag.com/issues/2014/12/3/renderings-sample-automatic-poem. Sometimes we also need to port a program to work on the Web, so that it can be accessible to readers.
Our objects of study, the things we are translating, are literary computer programs that we are working to bring into English from other languages.
Fascinating. I'm going to zoom in and ask you to define "poetry generator" for our readers.
Readers would do really well to follow the link I provided -- no definition would be as good as seeing one that a poet has written and that we worked to present in English on the Web. But a poetry generator is a computer program that (usually without much input from the user) outputs poetic texts.
I see, so the term is broad, and doesn't specify how, exactly, the generator works?
No, just as "poem" doesn't specify what the poetics behind a text are -- lyric, confessional, conceptual, traditional, experimental...
That's a great point. Though I do have a question about that. At the base of literature is a conveying of the human experience. We use words to share what’s inside of our skulls and put those thoughts in other people’s skulls. A poetry generator, in a way, hijacks that language, and removes all of the human input, the human intent. It seems your belief of literature is that its meaning is more contained in the interpretation than in the initial creation. Agree? Disagree?
I don't agree with any of your statement, starting with the premise (that literature can only be a communications vessel for sharing human experience). Where the statement goes from there also seems to me to be self-contradictory. A poetry generator has enough intentionality to be able to "hijack" something, but it lacks the intentionality that a human poem has.
The intentionality to "hijack" is with the creator, whereas the meaning behind the poem is through algorithm.
But you said "A poetry generator ... hijacks the language ..." To be brief, I think a valid purpose of poetry is to explore language and poetic forms.
I'm intrigued by how you mean when you first say you disagree with the premise.
This may have something to do with human experience, of course, but indirectly.
So literature you see as having multiple purposes: to share human experience, to explore language, and more. I can write a poem that explores the etymology of words, and through language get at something meaningful and interesting. So, why can't one write a program that explores language as well? That doesn't seem to be any sort of hijacking.
Some of the poetry generators and other text generators in Renderings make this point well: They are models of poetic practice in different languages and cultures. Or in one case, a model of how communist speeches were written and delivered. http://curamag.com/issues/2014/12/3/renderings-speeches
Very interesting. To boil an incredibly complex issue down to a short phrase, do you think a poetry generator can be creative in the same way a human can? Because language is a finite arsenal, they of course can be more "creative", but don't those (lack of boundaries) take something out of the game?
Well, I'm not trying to claim that the programs we translated in Renderings are creative in the sense that they themselves do creative work -- however, researchers in the Computational Creativity community are working to make poetry, visual art, story, music, and other generators that can strictly be called creative.
How is that?
So I would say it isn't necessary to have a creative system for it to be worthwhile, but computers can be creative.
First, one defines creativity in a formal way that allows non-creative systems to be distinguished from creative ones.
Then, one builds a computer system that is, according to this definition, creative.
This is an academic field similar to and overlapping with AI (Artificial Intelligence), and the idea is similar -- define intelligence very well, produce a system that exhibits intelligence.
We're talking macro, let's talk micro and explore a specific example. Can you tell me about Megawatt?
Sure, that's a program (and generated novel) that is based on Samuel Beckett's novel Watt. So, while some generators are based on genres or general ideas of text (communist speeches, Spanish lyrical poetry) some take a specific text as their basis.
Megawatt is also a deterministic program; it outputs only one thing. This isn't standard, but it's allowed!
How does the program take Watt and turn it into Megawatt? What's the mechanism?
It's sort of a compression system: The program is like a zipfile that expands to the text. The program doesn't "take Watt".
You'll have to excuse my novice language.
No, I just want to explain that you don't provide the novel as input or feed it into the system.
How is the novel involved?
I wrote the program, at first, to computationally produce passages from the novel.
Watt is an odd book, and some of the passages list every possible permutation of, say, four things. I sometimes say that computers had to be invented because Beckett wrote Watt (which he did during WWII).
So, for instance, if you find a text that includes "A! AA! AAA! AAAA! AAAAA! AAAAAA!". You could just type that out, as I just did. But you could also write a program that counts from 1 to 6 and produces that number of 'A's each time followed by an exclamation point. If you had 200 "AAA!"s you might find it much better to write the program. But in any case, you're building a model (using the computer) of how that text was composed.
I built a model (using the computer) of how several passages from Watt were composed -- afterwards, my program could write those passages using computation. And then, I absurdly extended them so that the already unreadable sections of Watt were even longer, but in exactly the same style as the original.
Ha! Have you read Megawatt?
I haven't read the entire text of the novel; I've only checked over it to see that the output is what I was looking for.
Not many authors can say that.
However, I consider writing the program Megawatt to be a very serious way of deeply reading Watt.
I'd agree. Would Beckett?
I think very few can say that they've engaged with a Beckett text in that particular way, too. I never met Beckett, and can't say what his position on this would be. I think there's a chance that it extends some of his humor, and if he agreed he might like that aspect.
Do you think a generated book would ever pass a lit-based Turing test? That is, will we ever be able to produce a novel (prose as opposed to poetry) that someone reads and believes was written by a human?
Oh, I should definitely explain why I don't care about this question.
I mean, as a computer scientist and someone interested in cognitive science I might, but as a poet and as someone who does creative work in digital media ... no. Some people want to imitate human writers with their computer systems. I don't mean extending Beckett to absurdity far beyond the original, which is not trying to fool you into thinking that the result is Beckett.
Of course, that's very different from what you're doing.
I mean producing a haiku that reads like a human haiku and can't be distinguished. When I encounter people who are doing this, I ask them to please stop! First, there is no shortage of haiku. There are already way too many of them. Second, why create a computer system that does exactly what people can already do in a way that's indistinguishable? I mean, again, as a scientific experiment it might make sense. As a literary project it does not.
Your second point implies any attempt at AI is irrelevant.
Well, it's more complex than that.
But my purpose in creating poetry generators and similar systems is not to be human-like, but to use the capabilities of the computer to explore language in new ways. Ways that the computer is good at. Simplest example: I write some programs that just continually produce output which scrolls by. A person can't do this -- not for a week, for instance. The person would keel over. But for a computer it's no problem. So, why not have poems appear like screensavers or like the water in fountains, so that we can look and read when we want to and look away at other times?
Again, that's the most basic sort of human/computer difference I can think of, I'm not sure it's a great illustration. Anyway, this doesn't mean that AI or machine learning or whatever sorts of techniques can't be used effectively in poetic systems. It's just that the goal is not "the imitation game" at least for me.
I have strong intuition you're going to vehemently reject this question, but do you think we should give prizes out for generated poetry and fiction the same way we do human-generated poetry and fiction? Could the two ever compete? Let me redact the second part of that question actually.
Heh. Actually I think it would be interesting for a computer system to win certain sorts of literary prizes for its output, from an institutional and attention-getting sort of standpoint.
But really it would make more sense to award the human creators prizes for the best systems.
How would we judge?
For one thing, the human creators will be more appreciative of the prizes.
Of course, at least so we can hang the medal on a fleshy neck.
I’d say any prize would have to be based on some kind of poetics, some idea of what is desirable ... I mean, you really a wide variety of different prizes for human literature. Best representing the views of the state, most in keeping with MFA program norms, most experimental in the way our group happens to like experimental work, etc.
Okay. Let's assume I'm personally dedicating $250,000 for the Nick Montfort Prize for Generated Poetry. You must choose five traits of judgment. What are they?
Engagement with and revelation about English (and/or other languages); engagement with poetic form; engagement with poetic traditions and techniques; compelling connection to specific authors, works, movements. Hm, and I think for a system there should be some value placed on conciseness of the program, understandability thereof, and status as free software. I think that's five (more or less). As painful as it is, you really need a committee to get criteria of this sort together. But that's an offhand answer.
Of course. We’re closing down on time so I’m going to stop us here and say if you’d like to buy Megawatt, you can at http://www.harvard.com/book/megawatt. I think if good interviews are marked by disagreement, then this one was great.
Ha ... well, it wouldn't be an interview if there weren't different views. You can also download the program Megawatt and generate your own identical novel with it: http://nickm.com/code/megawatt.py.
Awesome! Thanks for playing/putting up with me, Nick.
You bet, this was great.