Stanley and Me
[This is a talk I gave at Loyola University Chicago on November 8th, 2012.]
In January of this year, Stanley Fish published a series of online essays for The New York Times on digital humanities (this, then this, then this). To summarize: He doesn’t like it so much. And for those of us who like it a great deal, this was a curious mixture of good and bad news. It was bad news, because being publicly attacked by an extremely eminent public intellectual in the paper of record is a bit terrifying for scholars who already spend a good deal of time defending themselves in far less public venues. It was good news, though, because being attacked by a gray old man in the gray lady is exactly the sort of thing that tends to set our vast network in motion. Despite our much vaunted newness, the digital humanities community, in its online form, resembles nothing so much as the scholarly discourse of forty or fifty years ago. Back then, it was not uncommon to see attacks and parries playing out across several numbers of a journal, as so-and-so offered a response to Professor so-and-so’s late comments on Hopkins’s use of accentual meter. The Republic of Posts tends to do exactly this when dh-ers all around the world are reading the same thing on the same day — a phenomenon which happens very often online, and almost never off it. What’s more, Fish was naming names: friends of ours, books of ours, ideas we found compelling, and doctrines we held in common. Really, this was going to be ever so much fun.
But then he did something that ruined the whole thing for me in a stroke. In the midst of this mighty Oedipal struggle, he called me “perhaps the most sophisticated theorist of the burgeoning field,” and proceeded to comment on my work at some length and with a good deal of charity. He doesn’t like what I’m saying either, really, but — well, actually, it was hard to tell. He seemed to appreciate my work, or find it provocative…or something. Maybe he did like it?
“Damn it,” I thought, “What am I supposed to do now?” Suddenly, it seemed churlish to jump into the fray. I’ll admit to a certain guilty pleasure in having my work characterized by a famous scholar as “sophisticated” (even if I consider that statement absurd), but for the most part, I felt embarrassed by the whole thing. It would have been far easier to join those who had lately been skewered by the great Stanley Fish, because (as was obvious to everyone I know) they had all been treated unfairly. When this sort of thing happens, digital humanists tend to eschew the detached ironies of our more theory-obsessed forbears; our favorite rhetorical mode is something like righteous indignation tinged with self-deprecating humor. But I, having dodged a bullet in all of this, really had no cause for indignation of any kind. So in my flummoxed state, I decided to say nothing. I wrote a few tweets — mostly in response to a torrent of good-natured ribbing from far-flung colleagues and friends — and let the whole thing go on without me.
I’m already sounding a bit too grand and self-aborbed about all of this, so let me allay any fears you might have. I’m not about to give My Response to Stanley Fish. I’m not going provide some kind of witty (read, pompous) refutation of his arguments, defense of my friends’ arguments, or anything like that. I do, however, want to toss around a question that I have been thinking about for a long time, and which the Fish affair brought into sharp relief. Can you have computational text analysis and literary criticism at the same time? To get at that, we have to go back a bit.
Back when I started being a digital humanist — sometime in the mid-nineties — almost everyone I knew in the then-as-now “burgeoning” field was interested in the World Wide Web. “Hypertext” was the bold and exciting term on everyone’s lips, and there was much effusive talk of a new democratization of scholarship. I was as enthusiastic about that as everyone else, but I had managed to become entirely enthralled with a much older form of geekery. I had fallen in love with programming languages, and like a hammer looking for a nail, I was casting around for ways to use that in the study of literature. That led me to the quite considerable (and already decades old) body of scholarship surrounding what was then called “computational stylometry.”
It’s important to remember that back then, text analysis had nothing of the sheen is has now. I hope I won’t offend anyone who might have participated in this pioneering phase by saying that the field of text analysis was, at that time, a bit of a backwater, even within the barely-noticed field of humanities computing. Not many people were doing this sort of thing, and the ones who were tended to form a very narrow, if passionate band of enthusiasts. Nonetheless, I think I can fairly say that I devoured every article written on the subject from Busa onward, and was completely enchanted by the whole thing.
Still, I found myself experiencing the same moments of doubt that were even then being expressed among text analysis practitioners. Most disquieting for me were those moments in which stylometrists pointed, somewhat sheepishly, at the areas in which they felt their algorithms dare not tread. John Burrows and Hugh Craig, in one of the most stirring technical articles in the field at that time, had used principal component analysis to distinguish romantic from renaissance drama. After reading that article, I became convinced that such methods could lead to entirely new ways of thinking about periodization, gender, language, genre, cultural change, and much else. But in the midst of unfolding their marvelous mechanism, they casually averred that the sort of explanations offered by George Steiner — who had located the shift from renaissance to romantic drama in the loss of a “redemptive world-view” — were “well beyond the ambit of present computational stylistics” (Burrows 64).
I was mostly able to quiet my discomfort with this sort of disavowal. Perhaps it was only “beyond the ambit,” because the whole thing was still burgeoning? It struck me that there was basic research that needed to be done on text analytical methodologies, and much that needed to be theorized as well. Being a graduate student at the time, it even seemed possible that I was destined to boldly amble where no one had ambled before. There was a way to connect word frequency lists to redemptive world views; we just needed to figure it all out.
Then I read an attack on the work of Louis T. Milic by another scholar named Stanley E. Fish.
I don’t hear Milic referred to very often in digital humanities these days, but he was at least as important a pioneer as Roberto Busa or John Burrows. His 1967 book, entitled A Quantitative Approach to the Style of Jonathan Swift, is, as far as I can tell, the first book to tackle specifically literary critical questions using computers. I had encountered him many times in my attempt to read everything I could about computational methods in literary study — particularly since he was the author of over fifty scholarly articles. Milic died in 2003 (the year I got my PhD).
Fish’s attack was entitled, “What is Stylistics and Why Are They Saying Such Terrible Things About It?” (which is, of course, the second chapter of Fish’s highly influential 1980 book, Is There a Text in this Class?) I’ve never been exactly sure who was saying terrible things about stylistics at the time, but it seemed to me absolutely clear that no one was saying anything as terrible as what Fish was saying. Fish’s most recent attack on digital humanities is mild by comparison.
Milic’s work mostly involved word-frequency analysis; Fish was entirely interested in interpretation. So when Milic asks, toward the end of his study of Swift, “What interpretive inferences can be drawn from this material” (Milic qtd. in Fish 71), Fish is eager to hear the answer. And Milic’s answer is this: “The low frequency of initial determiners, taken together with the high frequency of initial connectives, makes [Swift] a writer who likes transitions and made much of connectives” (qtd. in Fish 71). Here’s what Fish had to say about that:
As the reader will no doubt have noticed, the two halves of this sentence present the same information in slightly different terms, even though its rhetoric suggests that something has been explained. […] There is, in short, no gain in understanding; the procedure has been executed, but it hasn’t gotten you anywhere. Stylisticians, however, are determined to get somewhere, and exactly where they are determined to get is indicated by Milic’s next sentence: “[Swift’s] use of series argues [that is, is a sign of or means] a fertile and well stocked mind.” Here, the procedure is not circular, but arbitrary. (72)
Milic, in other words, is made to look like a idiot (which, by the way, he most certainly was not). But when I read that passage, I found myself confronting an awkward truth about computational methods in literary study. For while Milic’s was an extreme example, it nonetheless brought forth an uncomfortable feeling I had about it. Text analysis very often consisted of the attempt to enlist sophisticated computational methods in the service of banality. Literary text analysis, in other words, was failing at its most essential task; it wasn’t generating literary criticism. The question for me became as explicit as this: Could we have a computational stylometry that made Stanley Fish happy?
A moment ago, I cited Burrows’s contention that George Steiner’s observations were “well beyond the ambit of present computational stylistics.” But we should pause to consider precisely what it is that lies beyond. Allow me to quote you some George Steiner (from his 1961 book The Death of Tragedy):
The tragic personage is broken by forces which can neither be fully understood nor overcome by rational means. […] Where the causes of disaster are temporal, where the conflict can be resolved through technical or social means, we may have serious drama, but not tragedy. More pliant divorce laws could not alter the fate of Agamemnon; social psychiatry is no answer to Oedipus. But saner economic relations or better plumbing can resolve some of the graver crises in the dramas of Ibsen. The distinction should be borne sharply in mind. Tragedy is irreparable. […] To ask of the gods why Oedipus should have been chosen for agony or why Macbeth should have met the Witches on his path, is to ask for reason and justification from the voiceless night. There is no answer. Why should there be? If there was, we would be dealing with just or unjust suffering, as do parables and cautionary tales, not with tragedy. And beyond the tragic, there lies no “happy ending” in some other dimension of place or time. The wounds are not healed and the broken spirit is not mended. In the norm of tragedy, there can be no compensation.
There are good reasons to ask, as the editor of Arts & Letters Daily recently did, whether George Steiner is a polyglot polymath or a eurocentric blowhard. But I will gladly say that whatever else Steiner is, he is a literary critic — someone who employs gorgeous writing in the service of profound ideas about human culture. If this is the thing that is “beyond the ambit” then why bother? Can any of us really say that we have seen this kind of thing — or anything remotely like it—arise from the quantitative analysis of literary texts? I’m quoting Steiner only because Burrows did; pick any literary or cultural critic you like and ask if you’ve ever seen an explication that arose from the workings of an algorithm get anywhere near it? I don’t mean algorithms generating beautiful and provocative critical prose; I mean algorithms leading the critic toward the sort of thing that literary critics want to say about human culture and its artifacts.
But even as Fish was exposing this rather embarrassing weakness, he was at the same time (somewhat inadvertently) showing us the way out. Fish’s real problem with computational stylometry in general (and Milic’s work in particular) is that nothing in the “machinery Milic cranks up” authorizes the leap from data to interpretation. And when nothing authorizes the leap, you can leap anywhere you like. Fish again:
One might conclude, for example, that Swift’s use of series argues the presence of the contiguity disorder described by Roman Jakobson in The Fundamentals of Language; or that Swift’s use of series argues an unwillingness to finish his sentences; or that Swift’s use of series argues an anal-retentive personality; or that Swift’s use of series argues a nominalist rather than a realist philosophy and is therefore evidence of a mind insufficiently stocked with abstract ideas (72).
My observation concerning that sarcastic list of interpretative possibilities is this: Every single one of those would be great. A text analytical argument that went in any of those directions would represent a momentous breakthrough for digital scholarship. Someone needs to write an article in which they use Levenstein minimal-edit distance and naive Bayesian inference to show that Jonathan Swift had an anal-retentive personality. That would be awesome.
The reason we don’t do that, of course, is because “nothing licenses” that kind of leap — just as Fish noted. To the degree that text analysis depends on scientific methods of measurement and evaluation, we might say the license at least feels more restrictive than usual. But what licenses Steiner’s leap? What licenses any leap? The text? Of course not. But why must “the data” adhere to a requirement that we long ago disavowed for text? Why can’t data function within the same interpretative regime — and fulfill the same hermeneutical functions — as text? Text sometimes operates in the service of objectivity and empiricism, but sometimes not (in the humanities, we might say almost never). Might the same be true of data? Is there a humanistic way to generate and understand data that licenses the kind of leaps we want to make?
Fish, in the end, was unable to imagine a hermeneutics that admitted neither of license nor constraint. I suspect he’s right about some system of license and constraint being necessary for meaning, though his notion of “interpretative communities” always struck as a bit too simple a description of what is undoubtedly a much more complex, multivariate phenomenon. (Wittgenstein, I suspect, gets us closer to a robust description of how meaning is constrained).
But let us return once again to beautiful literary criticism. Consider the following work by another very skilled literary critic:
Halfway through “Areopagitica” (1644), his celebration of freedom of publication, John Milton observes that the Presbyterian ministers who once complained of being censored by Episcopalian bishops have now become censors themselves. Indeed, he declares, when it comes to exercising a “tyranny over learning,” there is no difference between the two: “Bishops and Presbyters are the same to us both name and thing.” That is, not only are they acting similarly; their names are suspiciously alike.
In both names the prominent consonants are “b” and “p” and they form a chiasmic pattern: the initial consonant in “bishops” is “b”; “p” is the prominent consonant in the second syllable; the initial consonant in “presbyters” is “p” and “b” is strongly voiced at the beginning of the second syllable. The pattern of the consonants is the formal vehicle of the substantive argument, the argument that what is asserted to be different is really, if you look closely, the same. That argument is reinforced by the phonological fact that “b” and “p” are almost identical. Both are “bilabial plosives” (a class of only two members), sounds produced when the flow of air from the vocal tract is stopped by closing the lips.
[…] In the sentences that follow the declaration of equivalence, “b’s” and “p’s” proliferate in a veritable orgy of alliteration and consonance.
Even without the pointing provided by syntax, the dance of the “b’s” and “p’s” carries a message, and that message is made explicit when Milton reminds the presbyters that their own “late arguments …against the Prelats” should tell them that the effort to block free expression “meets for the most part with an event utterly opposite to the end which it drives at.” The stressed word in this climactic sentence is “opposite.” Can it be an accident that a word signifying difference has two “p’s” facing and mirroring each other across the weak divide of a syllable break? Opposite superficially, but internally, where it counts, the same.
I am quoting (and I hope you’ll forgive me for doing so at some length) from the famous broadside which occasioned these remarks: Stanley Fish’s piece in The New York Times.
Am I alone in thinking that this is lovely? I am aware of how problematic this is, given that Fish is using his interpretation of Milton to mount an argument for why digital humanists are deluded. Perhaps I should say that I like this kind of thing — a reading that takes us from the large to the small and back again. Meaning requires, among other things, an act of faith on the part of the listener, and in another context, I might well take the words of a famous Miltonist (like Fish) as evidence that he believes his own reading and thinks that I should believe it as well. I would hope to walk away with a deeper understanding of Milton, and of the theologico-orthographical currents of the seventeenth century.
There were lots of blog posts written in the wake of Fish’s article. The most famous of them, though, was Mark Liberman’s post entitled, The “Dance of the P’s and B’s”: Truth or Noise? which I can summarize as easily as I summarized Fish’s article at the start of this talk: Liberman thinks Fish’s reading is bullshit.
In the second part of Fish’s attack on stylometry, he says that “proponents of stylistics literally don’t know what they’re doing” (248), but that’s hardly a charge that will hold against Liberman. Liberman is able to establish beyond any doubt whatsoever that Fish’s “veritable orgy of consonance and alliteration” is, statistically speaking, more like a kiss on the cheek. After several paragraphs interspersed with graphs showing local distributions of bilabial plosivity, Liberman concludes that “a trivial application of statistical methods, humanistic or not, suggests that his idea is probably ‘false,’ ‘noise,’ and ‘mere play.’” He then asks, “Have I missed something?”
I don’t know how to answer that, exactly, but I am absolutely certain that I have missed something. Fish’s reading is literary criticism without data; Liberman’s is data without literary criticism. Neither one seems to me exactly right. I cannot imagine literary criticism — or indeed, any variety of humanistic discussion — flourishing under the constraint of having to prove its observations using statistical methods prior to advancing any claims. I also can’t imagine every attempt to use computational methods resulting in critical flights of world-historical significance.
But here’s one thing that I think makes either project impossible: the belief that the only hermeneutically legitimate procedure is one that moves from interpretative judgment to evidence. Here’s Fish:
I began with a substantive interpretive proposition — Milton believes that those who suffered under the tyrannical censorship of episcopal priests have turned into their oppressors despite apparent differences in worship and church structure — and, within the guiding light, indeed searchlight, of that proposition I noticed a pattern that could, I thought, be correlated with it. I then elaborated the correlation.
The direction of my inferences is critical: first the interpretive hypothesis and then the formal pattern, which attains the status of noticeability only because an interpretation already in place is picking it out.
The direction is the reverse in the digital humanities.
I said I wasn’t going to give my response to Fish, but here I must: This is complete and total nonsense. I hope I’ve made clear my admiration for Professor Fish’s work and its influence on my own thinking, because I cannot resist noting that on this issue, Fish literally doesn’t know what he’s talking about.
The idea that criticism only happens when you proceed from interpretative propositions to supportive patterns is, at best, a rhetorical dodge: “I had a very grand thought, and lo! I found it in the text!” To say that patterns cannot urge us toward interpretive propositions is to deny the text any serious role in the process of cognition. Has Fish never noticed a pattern and then had a thought about it? Is nothing perplexing in Milton? Is there nothing that reveals itself as a pattern, but without clear resolution or meaning? Interpretations undoubtedly define in advance the regime of our noticings (the elaboration of that idea is one of Fish’s greatest contributions to theoretical discourse), but surely that is not absolute. Otherwise, every interpretation would succeed by force of will (or force of narrow-mindedness). Surely the text’s ability to surprise us with noticed, but unresolved pattern is not only one of the great pleasures of the text, but the source of a great deal of its power.
I doubt very much if he’d disagree with my critique as stated, but if one concedes that reading is a process whereby patterns urge interpretations and interpretations bring forth patterns (a formulation that implies no natural sequence or direction), then we are halfway to saying why you might want to generate a computational program from a critical question or generate an interpretation from a computational procedure. Why not do both? In fact, why not do both within tight, concentric loops of reading, hacking, thinking, and interpreting?
In order to do this, we have to resist both of the apparent constraints that present themselves. A literary criticism that can only advance claims that are shown to be empirically valid is as deadening to the project of the humanities as a computational activity for which humanistic discourse lies permanently beyond its ambit. Risks need to be taken in both cases. In the former, the risk of saying something humanistically true but empirically false; in the latter case, of saying something empirically true that is humanistically false.
At the end of his essay, Fish states (with a certain resignation) that “whatever vision of the digital humanities is proclaimed, it will have little place for the likes of me and for the kind of criticism I practice.” I cannot really believe that that’s true, since my own thoughts on text analysis — which have very often been considered seditious by other text analysis practitioners — have their roots in another critic’s prescient, if brutal attacks on the very thing that has held me in thrall for nearly twenty years.
Stanley and me? We go way back.
Burrows, J. F. and D. H. Craig, “Lyrical Drama and the ‘Turbid Montebanks:’ Styles of Dialogue in Romantic and Renaissance Tragedy.” Computers and the Humanities 28 (1994): 63 – 86.
Fish, Stanley. Is There a Text in This Class? The Authority of Interpretive Communities. Cambridge, MA: Harvard UP, 1980.
Steiner, George. The Death of Tragedy. London: Faber, 1961.