Models and Meaning

Andrew Piper, Enumerations: Data and Literary Studies

1: Models

Andrew Piper's Enumerations: Data and Literary Studies, published by the University of Chicago Press last August, begins with a simple, if inflammatory, charge: that literary critics don't know how to generalize well. To prove his point, and to signal his ambition, Piper turns to the example of Erich Auerbach's Mimesis — an ironically singular, and single, selection. The revered book, as Piper reads it, makes claims about the entirety of "Western Literature."1 And yet it does so, Piper argues, on fallacious grounds. In each of Mimesis' chapters, a single passage stands in for an entire "great work." And these single "great works" stand in for "all of occidental literature." Piper calls this error an "epistemological tragedy": it suggests a basic incapacity to relate parts to wholes.2

For Piper, the problem persists. Literary critics still, like Auerbach, seek to generalize — we might, for example, make overarching claims about "the novel."3 And even those who resist generalization, Piper argues, must admit its importance. Even New Historicists like Catherine Gallagher and Stephen Greenblatt, who attend to the text's radical particularity, end up positing its enmeshments in general contexts. But for all our need to generalize, says Piper, we are still not quite up to the task. We cannot relate a text to its context with "anything like the rigor...applied to the individual document."4 What's called for, Piper argues, is a "science of generalization."5 In Enumerations, he supplies one: he argues that literary critics can use computational models to make better generalizations.

At first blush, this might seem a familiar story. For the past two decades, computational literary criticism has been strongly associated with the name of Franco Moretti, who, in his 2005 Distant Reading, made a similar argument. Moretti, like Piper, emphasized the need to make general claims — literary critics, he argued, could not understand the 0.5 percent of works that had been canonized unless they also looked at the other 99.5 percent. And like Piper, Moretti touted computation. There was simply no way, he wrote, for scholars to "close read" the "'great unread.'"6 By dint of sheer necessity, they must turn to "distant reading."

Piper, however, revises Moretti's line of thinking in two respects. First, rather than emphasize scale — or the need to examine more texts — he emphasizes representativeness. How, he asks, can we make truly representative generalizations, whether we're talking about single passages or entire genres? Second, rather than broadly suggesting using computational methods, he more specifically defends building "models." "Models," as Piper defines them, are computationally designed "representations" or "miniatures" of texts.7 They do, as is commonly said, "count words"; but they also perform more complex and diverse operations. By building them, Piper argues, we can better negotiate literary parts and wholes.

How might we go about that task? Ironically, it's hard to generalize. Throughout his broad-ranging book, Piper builds many different types of models — "vector space models," "topic models," and "predictive models," to name just three. And he uses them to answer many different types of questions. In his six chapters, he uses progressively more complex computational methods to address basic literary critical topics: punctuation, plot, topoi, fictionality, characterization, and corpus (or poetic career). His progress then culminates in self-consciousness. Piper turns his methods on himself and begins, somewhat whimsically, to build models that describe the book that we have just read. He gracefully admits, for example, that his chapters grow more confident as they proceed — that is, as he proceeded. For me, this final gesture of humility, even more than the book's bold, opening salvo, embodies its strengths. Enumerations offers two ambitious, theoretical defenses of modeling. But its case is ultimately most convincing for the degree to which, in practice, it is carefully circumscribed.

2: Words, words, words

Piper's first defense of modeling is born of a conviction: that word counts are more than just word counts. They are vital keys to the thing that we call a text's "meaning." This conviction finds support in the theory of "distributional semantics," which Piper takes from computational linguistics. According to the theory's "distributional hypothesis," the meaning of a word is tied to how often it appears in a certain context, and how often it appears, in that context, in relation to other words. The meaning of a document, by extension, is also tied to word frequencies and distributions. The theory has a certain intuitive force: one short cut to discovering that the words "peanut" and "almond" have similar meanings might be to note that they both often appear in noun phrases before the word "butter." It also has cognitive psychological backing: there's evidence that this is how our minds process meaning when we read.8

For literary critics, these concepts might not feel unfamiliar. Arguably, distributional semantics, with its emphasis on relations between words, encodes a structuralist theory of language. Piper ties it to poststructuralist theories, which also locate meaning "beyond the sentence." He cites Barthes' contention, in "The Death of the Author," that a text is "not a line of words releasing a single theological meaning," but rather a "multidimensional space in which a variety of writings blend and clash."9 He also mentions Deleuze's rhizomatic reading and Kristeva's intertextuality. What all of this is meant to suggest is that when we build models attuned to word frequencies, we capture something very close to linguistic meaning, not only as computational linguists conceive of it, but also as postructuralist theorists do.

Vector space models, in particular, are suited to the task. To explain how they work, Piper supplies three sentences from Goethe.

1. My dear friend, what a thing is the heart of man!
2. I treat my poor heart like a sick child.
3. I have possessed that heart, that noble soul.10

A vector space model, as Piper explains, converts each line into word frequencies. On that basis, it can measure their semantic relations. Here, it finds that the first two sentences are more similar to one another than they are to the third, because both contain "heart" and "a." Close readers may demur (and more on that later). But according to the theory of distributional semantics, at least, the model represents and compares these sentences' meanings.

Piper's appeal to distributional semantics is designed to preempt a common objection: that, when it comes to interpreting texts, word frequencies are functionally useless. Recently, Nan Z. Da has raised this objection anew. In her "the Computational Case against Computational Literary Studies," she emphatically repeats the point: that work like Piper's only deals with frequencies of words. "There are decisions made about which words or punctuations to count and decisions made about how to represent those counts. That is all."11 Such counting is "not interpretation," nor does it capture literary qualities, like "homonymy, figuration, polysemy, [and] irony."12 Piper, thanks to his investment in distributional semantics, repeats Da's formulation to opposite effect. Repetitions of words, he says, tell the "deep story" of literature; they are the "grooves and channels" of cultural expression.13

Piper's rejoinder is compelling. The distributional hypothesis's existence suggests that word counts cannot be dismissed, prima facie, as meaningless. But he will have to say more to convert skeptics to his cause. Even putting aside the concern about whether the distributional hypothesis is actually true, or congruent with postructuralist ideas, other questions emerge. Do other models, beyond vector space models, capture meaning as the distributional hypothesis construes it? And even if they do, does that entail that they can also capture the more specific literary qualities that Enumerations analyzes, like genre, character, or plot? Piper's appeal to the distributional hypothesis begins to answer an objection like Da's; but it does not settle the debate.

3: Charisma

Piper's second defense of modeling sounds a democratic note: that models help us argue not only more rigorously, but also more transparently. The classic close reader, Piper argues, delivers "imperious pronouncements" from on high, and relies on sheer "charisma" to convince. The model builder, by contrast, is beholden to a more regimented method; she can "define every step of [her] intellectual process." Computation might be "no stranger to opacity." "But done well and done openly," it can undo "the black box of critical charisma."14

The argument operates on multiple levels. On an intellectual level, it implies that model building is more accurate because it encourages self-consciousness — the model builder must check her intuitions against external mechanisms. On a political level, it suggests that model building might undo academic hierarchies. Piper, indeed, more broadly embraces that aim. In his afterword, he uses models to expose the operations of institutional privilege. Since 1970, he shows, 86 percent of all articles published in literary criticism's most prestigious journals have come from academics at the top 20% of PhD granting institutions.

But if Piper easily shows that model building can expose academic hierarchies, he is less convincing on the argument that it does not, itself, rely on charismatic authority. His case makes too much of close reading's reliance on charisma — close readers have to appeal to the text itself, after all, to check their intuitions and make persuasive arguments. And it makes too little of model-building's. In a world in which most literary critics still don't understand advanced statistics, even the most "open" of computational analyses will have to be taken on faith. Their appeals to advanced techne will likely still form the basis of their authority.

Literary critics are no strangers to ostentatious technics. New critics legitimated literary studies by producing the rigorous method that was "close reading." And since then we have applied our "theoretical lenses," our Jamesonian hermeneutics, our deconstructionist operations, and our readerly "methods," to the textual specimen. Distant reading is no freer of auratic technicity than prior critical practices, and, more likely, more enamored of it. In the course of analyzing character in his fourth chapter, for example, Piper quickly introduces a "character-feature tool," developed by himself and a colleague, "that tries to assess the practice of characterization according to 26 different dimensions," including "behavioral modality," "object-orientedness" and "a shifting suite of behavior, descriptive, and agential positions."15 It's hard to do anything but submit to the operations of this imposing and impenetrable tool.

The argument from charisma is, in any case, secondary. Whether it can come to play an important argumentative role will depend, first, on the truth of Piper's more essential claim: that models produce good textual interpretations. If they don't — if, for example, as Da argues, models do a poor job of capturing literary qualities — then the fact that they involve transparent processes will not be sufficient, on its own, to recommend them. Many processes can be laid out step by step - I can easily explain, for example, the steps by which I read only the last letter of every eight hundredth word of a literary text. But that does not mean that they should be embraced among our hermeneutic practices.

4: Bifocality

Fortunately for Piper, his argument can withstand scrutiny — at least on its own terms. Indeed, its resilience is often a function of its modesty. Models may or may not be the ultimate keys to all literary critical generalization. But they also don't have to be. Piper's more reserved critical practices make a good case for their targeted utility.

For one thing, Piper works with modeling not as one, but rather as one of many, critical methods. He explicitly advocates a practice by which close and distant reading inform one another, which he demonstrates in his first chapter. Here, Piper discusses punctuation in twentieth century poetry. Using a method referred to as "grep," he identifies a group of post-1900 poems that use particularly high numbers of periods per word; interestingly, he finds that they are disproportionately written by African American poets. Armed with that insight, Piper looks closely at Amiri Baraka and Angela Jackson's "high period poems" through a new kind of lens: why so many periods? And what might this have to do with the poems' racial politics? His close readings are compelling: they note that periods encode both struggle and potentiality, and function, like stab wounds, as repeated points of pain. Piper's readings are also critically generative: To my knowledge, there is as yet no study devoted to the poetics of excessive periods in African American poetry.

A major benefit of combining close and distant reading, in what Piper calls a "bifocal" process, is that it takes the pressure off distant reading to be comprehensive. Consider, again, Goethe's lines:

1. My dear friend, what a thing is the heart of man!
2. I treat my poor heart like a sick child.
3. I have possessed that heart, that noble soul.16

The vector space model, recall, demonstrates its power by judging that sentences one and two are most semantically similar. And yet, this example may at first appear to backfire. Why? Because if you are like me, you have an initial intuition that sentences two and three are actually most similar — their shared first-person pronoun is, in my reading, the overwhelming feature. But this doesn't mean that the model is wrong, or, for that matter, that I am; rather, the model compels me to consider another interpretation. It's true, after all — as Piper glosses his results — that, thanks to the repeated presence of "my" and "a" in sentences one and two, they commonly emphasize "possession" and "generality."17 When we start to consider models in this way, as partners in, rather than sole arbiters of, interpretation, it becomes hard to deny that they have something to contribute to our practices.

Even where Piper uses only distant reading, his examples do as much — if not more — to support his case than do his broader theoretical appeals. Consider his fourth chapter on fictionality, which asks the basic question: What is fiction? John Searle tells us that there is no essential difference between fictional and nonfictional language. And yet, Piper argues, are there not at least common features that might distinguish the two? To discover what they are, he uses a predictive, machine learning model, akin to that which distinguishes spam from non-spam email (see Dan's review for an explanation). The model, as it turns out, can distinguish fictional from nonfictional texts reliably, on the basis of many features. Piper uses these features to ground a more general, literary historical claim: that while we tend to think of modernism as having introduced an increased emphasis on characters' uncertain relations to the world — relations of "testing" or "hypothesis" — that emphasis actually belongs to fiction writ large.18

Da has this very type of model in mind when she fleshes out her argument that word counts do not meaningfully capture literary properties. Indeed, in one of the passages in which she makes this point, she refers to spam filters. When we use tools like spam filters, Da argues, we "[know] that human reading can catch more nuances, exceptions, ambiguities, and qualifications." But because these filters perform practical tasks, we don't care — "you wouldn't prefer to do it yourself." When we do literary criticism, however, we perform more than practical tasks, and so we should care — we should prefer to "do it ourselves." Da punctuates the argument by re-asserting a divide between spam filters' "practical" and literary criticism's "metaphorical" functions: "if we are employing tools whose express purpose is functional rather than metaphorical, then we must use them in accordance with their true functions."19

But just because a predictive model serves a "practical" function doesn't mean that it can't also generate "metaphorical" insight. Indeed, one might argue that predictive models work, practically, precisely because they pick up on metaphorically or descriptively relevant features. If spam emails, for example, can be identified, in part, because they reliably use words like "money," then maybe the monetary has something important to do with "spammyness." And if fiction can be reliably identified by its discussion of perception and the body, then maybe such "phenomenological" concerns — as Piper calls them — have something important to do with "fictionality." It may be true, as Da argues, that close reading — if it were possible at this scale — would pick up on other nuances. But it's not entirely clear that it would detect the same ones. The voracious close reader of a year's worth of novels might not alight on the fact that body words appear so frequently. And that is precisely Piper's point: that computers can inform, without overriding, our normal interpretive practices.

5: Conclusion; or, Having it Both Ways

Objections will, of course, remain. Some will argue that our mandate, as literary critics, is simply not to generalize — for critics in a New Historicist vein, as Piper notes, we exist primarily to attend to particularity. Others will argue that, whether we should generalize or not, to do so using computational methods is not worth the political cost— for critics like Sarah Brouillette, distant reading capitulates too much to the academy's neoliberalization. Piper's book will likely not sway members of either of those two groups. Rather than meticulously defend generalization, it largely presumes — though I think fairly — that we do at least sometimes want to generalize (are we really willing to invalidate every major book in our discipline that has made general claims?). And rather than defend modeling from the specific charge of neoliberalization, it endorses it on different political grounds: as a tool by which to undo academic hierarchies.

Enumerations is most likely to do persuasive work on critics who, like myself, are of two minds. Like many, I am open to the idea that models can contribute to our intellectual projects; but, like many, I cannot help but shudder at the thought of an English department devoted entirely to text mining. For those who share my feelings, Piper's book may appeal on these grounds: that it posits a critical world in which close and distant reading, aesthetic analysis and data analytics, cultural criticism and cultural analytics, can coexist. If Enumerations is any indication, then distant reading will not, and should not, overtake our changing discipline. We can have our computationally analyzed novel; and we can read it, too.

---

Tess McNulty is a PhD candidate at Harvard, working on contemporary literature and digital culture.

---

References