Virtual Poetics: A Thought-Experiment
Below is a video of a talk I gave at UC-Santa Barbara's "SyncDH" conference on May 8th, 2015-thanks to the incredible and innovative conference-organizing work of Ashley Champagne and Jeremy Douglass, who live-streamed the entire event and invited online participation.
The talk is on "virtual poetics." Basically, the talk is a first attempt at arguing that virtual literature-texts made by a computer-can productively reframe our understanding of actual literature.1 This project actually began tongue-in-cheek, producing the playful exercise of my last blog post on a "found" virtual poem. But the more I thought about virtual texts, the more seriously I found myself considering what they might have to tell us. In particular, I think they raise some interesting questions for the theory and methodology of DH.
I recommend watching the video for the story and visualizations, but I also wanted to try writing up and expanding the argument in prose form, which is what follows.
1. On virtual literature
What is a virtual text? Here, for example, is a virtual Spectator paragraph:
When being at St. James's Church. Damon, who so agreeably reconciles us to our selves the Happiness of such a Prodigy. He would have represented her in the following Passage related by Plutarch. That Author tells us, that the very next Day he commits a like Crime, and distinguishes him from himself by an huge Tub of Water, we shall be never able to make any Judgment of the Performance; every thing disposes its self into Order as it were frozen and congealed by Winter, are now turned loose, and set forth the Vanity of the Sex following one another in Countenance; and one could say nothing more to it at present .
A computer (really, a third-order Markov chain) wrote this-heavily influenced, you might say, by Addison and Steele. This was its working method: beforehand, it had read all 635 essays of the 1711-12 periodical, keeping track of which words appear after every unique three-word sequence. So for example, after the three words (or tokens) "St.", "James", and "'s", the following words appear in the real _Spectator _essays: "Coffee-House," "Park," "Street," "Palace," "Church", the word "will", a comma, and a period. The computer wrote this exquisite Augustan prose by simply picking a word from this list randomly, whenever the last three words it had written were "St. James 's." Then, if it picked "Church," and its last three words became "James 's Church," it would pick from their list of subsequent words. And so on. The colors reflect the number of options from which each word was chosen: black meaning there was only a single option (given the previous three words), and the scale from dark blue to bright purple reflecting the range from two to many options.
This text is not only "virtual" in the sense that it was written by an algorithm. It's also virtual in another, more interesting way. The text is simultaneously unreal (it never appears in The Spectator) and real (it was constructed from The Spectator). In fact, the colors reveal how it dips in and out of virtuality: while the stretches of black words indicate stretches of real passages from The Spectator, the blue and purple words indicate moments that the virtual text "turns" to explore a path of language made _possible _by The Spectator, but which was never before actualized. These "turns" are often hilarious: never before was someone, for Addison, distinguished by "an huge Tub of Water." But also, arguably, they indicate something about the generative logic of Addison's prose. Phrases like "Vanity of the" and "Judgment of the", or "in the following" and "as it were", are followed, in The Spectator, by a wide diversity of other words. They provide not only the virtual Addison, but also arguably the _real _Addison, with a garden of forking paths of language to explore-with moments at which the potentiality of language is amplified in a way useful for the production of more real language.2
If, for Deleuze, the virtual is the way in which something differs from itself, then perhaps we can see this virtual Spectator passage as visualizing the ways in which Addison's prose can, at certain moments more easily than others, differ from itself-by turning on itself, reconstituting itself, repositioning itself within its own discursive field. I think this is one powerful lesson virtual textuality has to teach us about actual textuality: that a given literary text is not uniformly actual, or uniformly determined, but instead constituted by and through a range of determinants and possibilities-a range made visible through the study of virtual textuality.
2. Lens vs. Cause
But this post isn't primarily meant as a philosophical reflection on textuality. Perhaps more importantly, I think the study of virtual texts bears consequences for the theory and practice of digital humanities. As Paul Feyerabend, the philosopher of science, wrote in Against Method, "we need a dream-world in order to discover the features of the real world we think we inhabit (and which may actually be just another dream-world)."3 The study of virtual literature, even and maybe especially in the context of distant reading, could usefully provide a dream-world referent for our study of actual literature, destabilizing our implicit assumptions about DH epistemology by playfully leading to what Feyerabend calls a "new conceptual system that suspends, or clashes with, the most carefully established observational results, the most plausible theoretical principles, and introduces perceptions that cannot form part of the existing perceptual world." Indeed, perhaps the quasi-scientific method inherent in the project of distant reading requires, for its fulfillment, its own contradictory epistemology-one akin to the creative, "'pataphysical" practices invoked by Stephen Ramsay and which lay at the root of experiments in textuality conducted by members of the Oulipo.4
In particular, I think virtual literature helps raise specific questions about established modes of interpretation in DH research. One such question is the following.
I wonder if DH studies of actual texts-including my own as much as anyone else's!-could be said sometimes to encounter a particular interpretive problem that I've nick-named the Lens & Cause Problem. I think this problem is occasioned by a certain opacity to meaning within numerical data. By definition, quantitative methods can only count "countable" things-that is, things that repeat. DH studies can't "count" texts themselves; rather, they count smaller, repeatable things _in _texts, like words. But then, let's say we've discovered that 11% of the words in Moby-Dick are verbs: what does this tell us? Not much. Is 11% a lot, or not that much? Well, in comparison to what? This is where DH studies typically turn to metadata (author, genre, publication date) to provide an interpretive lens through which a mass of numerical data becomes meaningfully differentiated. Melville's 11% is the highest of American authors in the 1850s, let's say [note: this is a totally hypothetical statistic]. Suddenly our interpretive wheels start turning: is that owing to the sea-voyage plot? To the rough adventurousness of that novel? To Melville's aggressive romanticism?
But herein lies the problem, I think. When we speculate in this way, by asking what about Melville as an author "leads to" his greater usage of verbs, has the _lens _through which that data became meaningfully differentiated (i.e. Melville vs. other authors) transformed, almost invisibly, to become at the same time the cause of that differentiation? In other words, if metadata serves first as a lens for the data, leading to a descriptive statistic about a given author or genre in comparison to others, it can then sometimes seem to quietly reify itself into its own explanation. Melville, quietly becomes Melville-ness.
There's a politics to this problem as well. Imagine a (hypothetical) neuroscience study of gender and the brain: data is gathered from male- and female-bodied brains. Gender is the metadata by which that data is meaningfully differentiated: let's say women have more activity in a certain region of the brain. How easy would it then be to speculate that this activity, made visible by gender, is at the same time _caused _by gender? Former Harvard president Larry Summers once argued that since men score better than women on mathematical tests (i.e. gender as lens), there must be something inherent to woman-ness that explains why fewer women occupy math and science positions at Harvard University (i.e. gender as cause)-and then he was fired. For good reason.
What's missing in these accounts, both in gender-psychology and DH, is a more robust explanation of inter-mediating and confounding causal factors. Is it female-_ness _that causes women to score more poorly on math tests, or is this owing instead to discrimination, social conditioning, or asymmetrical expectations in the U.S. educational system? Is it really Melville-_ness _that explains his 11% of verbs, or is this owing instead to any number of a wide range of inter-mediating or confounding factors? Confounding factors could include, for instance, generic rather than authorial differences: are the other authors included in the comparison writing domestic or sentimental fiction?-in which case, perhaps the relevant explanatory framework is of those genres, rather than of Melville as a distinctive authorial personality. Such confounding factors are especially problematic when collections of texts are made to represent not an author, but a subtler concept, like (as a hypothetical example) "novels with free-indirect discourse (FID) in them" versus "novels without FID." Although this difference in FID is true about the two sets of novels, how many other differences could also be true about them? The FID novels might happen to be historically distributed differently; they might happen to have different generic contexts; they might happen to contain different authorial idiosyncrasies; and so on. If, let's say, we found that the set of "FID novels" had simpler diction than the other set-given these potential confounding factors, could we really then claim that FID tends to _produce _simpler diction?
Even more interesting for the question of virtual textuality in DH, however, is the possibility of "inter-mediating" factors that could trouble the connection between, say, verbs and a given author or genre. We know, for instance, that verbs negatively correlate with adjectives and nouns. We also know that the number of verbs would be directly related to patterns of sentence length and syntax. In fact, a whole range of aspects of textuality (many of which we probably don't yet know about) exerts powerful intra-textual forces upon the one currently under the microscope-forces that could, and should, destabilize any interpretation that holds up a single such textual aspect as the hallmark of a given author, genre, or any other metadata category. Is it a hallmark, or an epiphenomenon?
Finally, not only is there an abstractly epistemological (and political) aspect to the Lens & Cause Problem. I wonder if there is a more concrete, disciplinary dimension to it, as well: that is, whether it could also help explain why DH studies frequently don't tell us as many things (as we would like) that we didn't already know. By relying on metadata (which usually registers a widely accepted conceptual category) to make sense of their data, DH studies enable quantitative meaning while at the same time foreclosing the possibility that that meaning could cut across-or extend outside of-the very concepts that gave rise to it.
3. Virtual epistemology
Because of all of these thorny interpretive difficulties, a _virtual _DH methodology-a "virtual poetics"-becomes especially interesting. Crucially, such a "dream-world" methodology seems to side-step many of them. This is because it allows us to compare an actual text not simply with other actual texts, where the "lens" of metadata (author, genre, year, etc.) exerts a heavy hand in any eventual interpretation we make of the data. Instead, virtual poetics allows us to compare a single actual text with-_itself. _As we saw above with the Spectator, a single text can generate any number of virtual permutations of itself. Moreover, the generative logic behind those permutations can be formulated and modified precisely. How would the real Spectator compare to a virtual one, for instance, that prohibited adjectives, or that required an abstract noun per sentence?
To ask this question-and this is what the recorded talk above focuses on-I conducted a little experiment comparing Shakespeare's actual 154 sonnets to 20,000 _virtual _Shakespeare sonnets. These virtual poems, generated from a second-order Markov chain trained on the real sonnets, all rhyme ababcdcdefefgg, and all have exactly ten syllables per line. But, although they share with the real sonnets these _formal _determinants (rhyme and line length), they don't share other determinants involved when Shakespeare wrote his sonnets: forces like semantic coherence and aesthetic interest. If we think of form and force as two sides of a coin, then without the forces of semantics and aesthetics, how do these virtual forms differ from their actual counterparts? To return to the brain metaphor: rather than comparing Shakespeare's brain to other brains, here, we've removed parts of Shakespeare's brain (semantics and aesthetics), and then compared its behavior with the original one. Here's an example virtual Shakespeare sonnet:
"Prayers" and "affairs"-not a bad rhyme, right? But, we can already see the absence of certain determinants involved when Shakespeare actually wrote his sonnets: he would never rhyme "I" and "thy," that is, allow such a radical form of enjambment in which a line ended on a possessive determiner ("thy"). To take a more distant reading approach, we can plot Shakespeare's actual and virtual sonnets on an identical morpho-space, one that I along with others in the Literary Lab (Mark Algee-Hewitt, Jonathan Sensenbaugh, J.D. Porter, and Justin Tackett) have constructed in our project on the history of poetic form. This morpho-space successfully disaggregates poems into those predominantly in iambic, trochaic, anapestic, or dactylic meter:
Shakespeare's real sonnets (in blue) cluster in the iambic quadrant in the top left: their lines are predominantly in a head-final or rising rhythm (high on Y-axis); and they have relatively few ternary feet (to the left on the X-axis). By contrast, the virtual poems are split between the two left quadrants, or between iambic and trochaic verse; but, they're as loath as Shakespeare was to include ternary feet. So, although a second-order Markov chain (in which any two, but not three, consecutive words from the virtual sonnets appear somewhere in the actual sonnets) is sufficient to replicate the _binary _quality of Shakespeare's iambic pentameter, it isn't sufficient to replicate its _head-final _quality (i.e. what separates iambic from trochaic verse). This means that some other force was involved in making Shakespeare's ten-syllable lines iambic-an additional force of formal patterning not captured at the level of bigrams from which the Markov model is working. Many critics have conjectured that English is naturally iambic; but, even when we limit the virtual poems' vocabulary to Shakespeare's, they do not naturally fall into an iambic meter.
Of course, it isn't all that surprising that more was involved in crafting Shakespeare's iambic pentameter than what was captured by this simple Markov model. But what I find most interesting about this experiment is not its result, but its method and its potential. Rather than comparing Shakespeare's sonnets to others from the period or elsewhere, here we've compared them to a virtual and precise deformation of themselves. What's striking about this method is the way in which it allows us to investigate what's involved, intra-textually, in producing the formal effects we've quantified from the sonnets. Which aspects of Shakespeare's sonnet form are "necessary," given the general stress-system of English words or Shakespeare's diction, and which derive from additional forces of artifice? Virtual textuality thus allows us to find the ways in which, and the degree to which, a literary text is able to differ from itself-a unique approach to understanding textual specificity. Thus, as the colors in the virtual _Spectator _passage above visualized, virtual textuality makes visible the _modality _of literary form, its own internal and dynamic range from necessity to freedom.
I am certainly not the first to argue that virtualizations of literature are useful for its study. Graham Sack has been working for years on applying techniques of generative modeling to literary structures such as the character-network; and I know Michael Gavin is also doing some interesting work using agent-based modeling. However, I do think there are interesting differences-at least in emphasis and theoretical purchase-between concepts of modeling and concepts of virtuality, especially in terms of virtual textuality.
In fact, as Marissa Gemma's research shows, corpus linguistics has a term for such frequent, generative, and discourse-specific phrases, like "the ways in which" in academic prose: lexical bundles. Gemma and I are currently collaborating on a project on the history and function of lexical bundles in eighteenth and nineteenth-century British and American fiction.
Paul Feyerabend, Against Method (London: Verso, 1993), 22.
Stephen Ramsay, Reading Machines: Toward an Algorithmic Criticism (Urbana: U of Illinois, 2011), 18-31.