Digital Evidence and Literary Change
Digital Evidence and Literary Change
“Distant Horizons is of compelling interest to digital humanists. But its true audience is a wider society of literary and other humanities scholars spanning across fields, periods, approaches, and levels. For this larger audience, Ted Underwood goes out of his way to make distant reading accessible, inviting, and persuasive. This innovative book is the breakout work digital humanists have been waiting for, and it is positioned to be a landmark work in literary scholarship at large.”
Alan Liu, author of Friending the Past: The Sense of History in the Digital Age
“Distant Horizons not only proves that Ted Underwood is defining the field of cultural analytics as it emerges; it shows us why. Combining literary theory with a deep understanding of computational methods, this volume demonstrates and effectively argues that quantitative analysis is best used not to find objective truths but to explore perspectives, both historically local and theoretical. It is at once a primer for quantitative literacy and a historically sensitive exploration of gender, genre, character, and audience, putting paid once and for all to the notion that statistical methods have no place in hermeneutics.”
Laura Mandell, author of Breaking the Book: Print Humanities in the Digital Age
"This is a substantive contribution to the debate over what Franco Moretti dubbed 'distant reading' and its place in the study of literature. Underwood engages contemporary scholarship, building and testing hypotheses based in the last 20 years of work. . . . Though technical in method, the book is engaging, and Underwood punctuates the argument with data-rich graphs and tables. The volume concludes with a healthy, skeptical consideration of the dangers of distant reading that nevertheless argues for the place of digital reading, alongside more traditional literary inquiry, as a tool for 'learning to doubt one's own perspective.'"
“Ted Underwood’s Distant Horizons: Digital Evidence and Literary Change is an insightful account of the applications of computational methods in literary studies and a striking demonstration of the labor that goes into making them work. The book digests and draws implications from the more technical studies that Underwood has conducted during the past decade
with collaborators in literature and computer science, putting them together into a coherent account of an analytical approach with which most humanists are only generally familiar. Along the way, it also offers thought-provoking accounts of periodization, genre, theme, and gender in modern English literature based on Underwood’s computational research.“
Daniel Rosenberg | Modern Philology
"Underwood’s knowledge of his own materials and methods and his ability to explain them to uninitiated readers are truly exceptional, while he is admirably open to changing his mind and curious to try new approaches. His judicious application of computational tools to digitized corpora of modern printed texts is rightly influential, and I will continue to follow his innovations with pleasure and interest."
Table of Contents
Preface: The Curve of the Literary Horizon
2 The Life Spans of Genres
3 The Long Arc of Prestige
4 Metamorphoses of Gender
5 The Risks of Distant Reading
Appendix A: Data
Appendix B: Methods
This is a book about recent discoveries in literary history. The word discovery may sound odd, because the things that matter in literary history are usually arguments, not discoveries. Although lost manuscripts do occasionally turn up in an attic, uncovering new evidence is rarely the main purpose of literary research. Instead, scholars reinterpret the well-known outlines of the past (Romantic, Victorian, modern) by drawing new connections be- tween texts or by moving something marginal to center stage.
Or so I thought ten years ago. Over the past decade, I have gradually lost confidence that the broad outlines of the literary past are as well known as I once thought. As scholars have learned to compare thousands of volumes at a time, we have stumbled onto broad, century-spanning trends that are not described in textbooks and not explained by period concepts. It is becoming clear that we have narrated literary history as a sequence of discrete movements and periods because chunks of that size are about as much of the past as a single person could remember and discuss at one time. Apparently, longer arcs of change have been hidden from us by their sheer scale—just as you can drive across a continent noticing mountains and political boundaries but never the curvature of the earth. A single pair of eyes at ground level can’t grasp the curve of the horizon, and arguments limited by a single reader’s memory can’t reveal the largest patterns organizing literary history.
In this book, I explore some of those patterns and explain how new approaches to literary research are making them visible. I follow in the footsteps of many other scholars who have posed broad social questions about literature. The work described here owes something to twentieth-century projects like book history, stylistics, and the sociology of literature, as well as to the more recent fusion of those projects that goes under Franco Moretti’s term “distant reading.” I will be less concerned to trace academic genealogies than to describe specific discoveries that are redrawing our map of the last three hundred years of English-language literature. The first four chapters are each organized around a different historical discovery—illuminating literary language, genre, aesthetic judgment, and the history of gender.
The book also describes the new methods required for large-scale research and discusses the reservations many people feel about applying computers to literature. But I do not approach those questions as they are commonly framed—as a struggle that pits critical tradition against a new technological initiative called “digital humanities.” That frame has been popular for several reasons. It fits a familiar narrative that casts digital computers as the main agents of change in recent history, as well as an even older narrative organized around conflict between machines and culture. Well-worn stories of that kind come with a familiar set of moral coordinates, making it easy for observers to express an opinion about changes labeled “digital” without studying the changes themselves in much detail.
That’s unfortunate, because the advances that have made large historical patterns visible have less to do with computers than with new ideas about modeling and interpretation. Computers themselves, after all, are not very new; scholars have been applying them to literary language for more than fifty years. If digital technology had been the only thing required for a new approach to literary history, this book would have appeared long ago. But in the 1970s, the application of computers to literature often produced arguments about sentence length or about Jonathan Swift’s favorite words. Most scholars doubted that a computer’s ability to precisely measure those linguistic details would, in it- self, transform the history of literary pleasure. In my view, they were right to be doubtful. As Stanley Fish pointed out, it’s one thing to prove that Swift uses a lot of connective words and another to give that isolated fact a literary interpretation.
So what changed over the last fifty years? Admittedly, scale is one part of the story. Up through the 1980s, quantitative exploration of literary history tended to be founded on relatively small collections, often focused on individual authors. The expansion of digital libraries has made it easier to pose broad historical questions, and historical breadth has given quantitative inquiry a better social foundation. (This book, for instance, is deeply indebted to HathiTrust Digital Library, among other sources.) But sheer scale is only part of the story. The discoveries described in this book do depend on a wide field of view—as the curve of the horizon only becomes visible some distance above the earth. But a wide field of view is not enough, by itself, to give linguistic details a literary meaning.
Numbers are becoming more useful in literary study for reasons that are theoretical rather than technical. It is not that computers got faster or disks got bigger but that we have recently graduated from measuring variables to framing models of literary concepts. Since a model defines a relationship between variables, a mode of inquiry founded on models can study relationships rather than isolated facts. Instead of starting with, say, the frequency of connective words, quantitative literary research now starts with social evidence about things that really interest readers of literature—like audience, genre, character, and gender. The literary meaning of those phenomena comes, in a familiar way, from historically grounded interpretive communities. Numbers enter the picture not as an objective foundation for meaning somewhere outside history but as a way to establish comparative relationships between different parts of the historical record.
I realize this is a loosely sketched picture. The word model itself is not yet common in literary study, so chapter 1 will spend some time explaining what it means to frame a statistical model (especially a “predictive model”) of a literary concept. All I want to say at the outset is that the advances making this book possible were not mostly a matter of computing power. They have depended instead on a debate about modeling, learning, and interpretation that is currently transforming fields from statistics to psychology. I will dip into that debate throughout the book, and survey it more fully in an appendix on “Methods,” in order to give readers a glimpse of important developments in recent intellectual history. But in the end, this is a book about the history of English literature, focusing especially on Anglo-American writers. Instead of emphasizing new methods, I will underline specific literary insights they make possible. Each chapter will be organized as a historical argument.
The first chapter suggests that many well-known changes in eighteenth-, nineteenth-, and twentieth-century fiction can be understood as parts of a single differentiating process that de- fined the subject, style, and pace of fiction through opposition to nonfiction. We already know about parts of this story. Scholars of eighteenth-century fiction have discussed the end of feigned autobiography, scholars of the nineteenth century talk about emphasis on visual detail, and scholars of modernism discuss the decline of the omniscient narrator. From time to time, more controversially, a critic will suggest that some of these changes could be unified under the banner of a broader shift from “telling” toward “showing.” But it has been difficult to make a unified story persuasive: Victorians and postmodernists, for instance, may refuse to line up with modernist triumphalism about the rise of the impersonal, limited narrator. With the broader perspective made possible by quantitative evidence, it is now possible to see all these changes as stages of a long differentiating process. A wide range of artistic movements, often said to conflict with each other, sometimes said to have sought rapprochement with “ordinary language,” have all actually pushed fiction farther away from the language, themes, and narrative strategies of nonfiction.
Of course, “fiction” is a rather broad genre; literary scholars are more commonly interested in the history of subgenres like Gothic or detective fiction. Chapter 2 zooms in on those concepts in order to explain how new methods can support a perspectival approach to genre. The Aristotelian conception of genres as natural literary kinds has given way over the last fifty years to a warier approach that treats genres as historically contingent institutions. Instead of trying to give science fiction a stable definition, critics increasingly propose that it is, at bottom, just the loose grouping of works that different historical actors have called “science fiction.” This implies that science fiction may have meant different things at different times and puts critics who want to talk about science fiction before the 1920s in an awkward position, since none of those literary traditions were called science fiction by their original readers.
Genres are not the only human creations that change their meanings with time. The interpretive problems that confront a history of genre are rooted in the perspectival dimension of history itself, and they run too deep to be solved neatly. The meaning of a term like science fiction will always depend on an observer’s location. But one of the central arguments of this book is that contemporary quantitative methods can be very good at representing perspectival problems and can give us leverage on that dimension of history.
Questions of perspective may be the last place we would expect to encounter math. In the twentieth century, numbers were used mostly for physical measurements (or demographic counts) that didn’t vary greatly from one observer to another. Those associations have given many people the impression that Arabic numbers are somehow in themselves objective or aspire to be independent of social context. But if we look with fresh eyes at contemporary quantitative methods, we may notice that they are not distinguished by any aspiration to objectivity. Machine learning, in particular, is causing public scandal because it tends to be all too sensitive to subjective contexts.
When scholars explicitly define a concept, we can craft a definition that aspires to neutrality. But the models produced by machine learning don’t rely on explicit definitions; instead, they learn concepts entirely from illustrative examples. Learning from examples makes machine learning flexible but also very apt to pick up the assumptions or prejudices latent in a particular selection of evidence. This has become a huge problem for institutions that are expected to be neutral arbiters. We don’t want a bank’s judgments about creditworthiness to be shaped by assumptions about gender or race. But a model that learns about creditworthiness from examples of approved and rejected loans is very likely to absorb the biases of the people who approved or rejected them. Institutions that strive to be unbiased might well choose to avoid machine learning. When we’re reasoning about the past, on the other hand, our aim is usually to acknowledge and explore biases, not to efface them. Understanding the subjective preferences implicit in a particular selection of literary works, for instance, may be exactly the goal of our research. For this kind of project, it is not a problem but a positive advantage that machine learning tends to absorb assumptions latent in the evidence it is trained on. By training models on evidence selected by different people, we can crystallize different social perspectives and compare them rigorously to each other.
This approach, which I call “perspectival modeling,” has taken shape only in the last few years. Readers who are familiar with other ways of using machine learning may need to set some assumptions aside. The models created in this book are supervised: that is, they always start from evidence labeled by human readers. But unlike supervised models that try to divine the real author of an anonymous text, perspectival models do not aim simply to reproduce human judgment. They are used instead to measure the parallax between different observers.
This strategy will have many applications in the pages that follow. The second chapter, for instance, uses it to pose questions about the history of genre. In some cases, genres defined by observers in different periods turn out to align better than their names might suggest. A model trained on nineteenth-century “scientific romance” finds it easy to recognize contemporary “science fiction” as a version of the same thing. In other cases, different perspectives turn out to be incompatible: the various traditions readers have called “Gothic,” for instance, aren’t well recognized by a single model. Evidence like this can help historians move beyond sterile arguments about lumping and splitting and toward a more flexible debate that acknowledges boundaries with different degrees of blurriness.
Chapter 3 begins to explain how questions of form and genre intersect with grittier aspects of literary production and distribution. This requires enriching a library of texts with social context—so we know, for instance, which works became commercial successes or critical favorites. Armed with that evidence, scholars can ask how literary trends were related to the pressures exerted by the marketplace or by changing patterns of critical judgment. This inquiry reveals a strikingly regular pattern, where the criteria defining literary prominence align strongly with directions of change across long periods of time. The arc of literary history is long, but it bends (so to speak) toward prestige. At this point, we are no longer simply transforming familiar accounts of history by backing up to take a longer view of them. If standards of aesthetic judgment have remained relatively stable for centuries at a time, and have shaped literary change over equally long timelines, then we are looking at an account of literary history that is basically at odds with the story of rapid generational reversal told in our textbooks and anthologies.
The first three chapters of the book describe dimensions of literary history (like reception and genre) where volumes can be discussed for the most part as wholes. Topics like plot and character are harder to trace across long timelines because they require divisions below the volume level that are challenging to tease out algorithmically. But with collaborative support from computer scientists, it is also possible to make some progress on those topics. Chapter 4 explores the history of characterization, looking in particular at the way fictional characters are shaped by implicit assumptions about gender. Once again, perspectival models provide crucial leverage for my argument. For instance, one way to ask how strongly characterization has been gendered is to ask how easy it would be to distinguish fictional women from men, using only the things they are represented as doing in the text. When first names and pronouns are set aside, can a model still predict a character’s grammatical gender? And if so, how do perspectives on gender vary across time? Using tools built in part by David Bamman, I have been able to show that the implicit gendering of character grows steadily blurrier from 1840 to the present. More interesting, of course, are the specific details that signify gender. These are not always obvious: in the middle of the twentieth century, it suddenly becomes feminine to smile but masculine to grin. Perhaps most interesting of all: the details that predict a character’s gender turn out to be extremely volatile. Fictive gender is not the same thing today that it was in 1840. Along the way, we’ll stumble over some counterintuitive trends in the social history of authorship—notably, a 50% decline in the fraction of English-language fiction written by women between 1850 and 1970.
The approach to literary history I have outlined above is controversial, to say the least. Literary arguments don’t ordinarily use numbers, and many scholars doubt that numbers can ever play an important role in the humanities. The fifth chapter of this book responds to those concerns in depth.
I delay this controversy to the end of the book because I don’t see it as a struggle between competing philosophies that could be decided in advance by invoking first principles. Doubts about the value of large-scale quantitative research are doubts about the inherent interest of a new perspective on the past, and there is simply no way to know whether a new perspective will be interesting until you have explored it. At the end of this book, after exploring a new scale of description, I will weigh its inherent interest against the price humanists might have to pay for this expansion of their horizons.
There is, to be sure, a price to be paid for all knowledge. But in this case, the price is institutional rather than philosophical. We are not looking at a debate like the struggle between structuralism and poststructuralism, where one perspective had to be abandoned in order to adopt another. Distant reading is simply a new scale of description. It doesn’t conflict with close reading any more than an anatomical diagram of your hand would conflict with the chemical reactions going on inside your cells. Instead of displacing previous scales of literary description, distant reading has the potential to expand the discipline—rather as biochemistry expanded chemistry toward a larger scale of analysis. And yet there is admittedly a cost, even to expansion: new kinds of training could stretch scholars and perhaps change the character of a literature department. So in the fifth chapter, I meditate on the temperament and training required for quantitative research in the humanities and let readers decide whether the new perspective unfolded in the first four chapters would be worth paying the associated price.
But costs can only be weighed against benefits after we see what long timelines reveal. The most I can achieve in a preface is to clear up a few misunderstandings that might scare readers away at the outset. One concern, in particular, may spring to mind the moment you open this book and see a graph: that quantitative methods seek to strip away the interpretive dimension of the humanities in order to produce objective knowledge. This notion springs, I think, from a failure of communication between humanists and scientists. To make a long story short: numbers are not inherently more or less objective than words. Numbers are just signs created by human beings to help us reason about questions of degree. Like other arguments about the past, a statistical model is a tentative interpretation of evidence. Expressing a model mathematically has the advantage of making some assumptions explicit (including, especially, assumptions about quantity and degree). But numbers have no special power to settle questions: assumptions and inferences still have to be hammered out through a familiar process of debate. In literary history, moreover, scholars will often be using statistics to model aspects of the world that are themselves subjective beliefs. In exploring genre, for instance, I have modeled variables like “the probability that a particular group of observers in 1973 would have thought this was an example of detective fiction.”
In other words, a quantitative approach to literature does not have to be premised on a belief that literary history is governed by any drily factual Marxist or Darwinian logic. While this book sketches patterns of change across long timelines, it will generally resist the assumption that literary history can be explained by a familiar master narrative. The first chapter of this book will begin where readers of literature usually begin—by exploring the details of two particular stories. As we back up, it is true, those details will start to organize themselves into larger patterns shared by many other books. And in an attempt to understand those patterns, we will start to form generalizations we call models. Those models will reveal large patterns that scholars have previously failed to describe. But they won’t eliminate perspectival differences and debate. Quantitative models are no more objective than any other historical interpretation; they are just another way to grapple with the mystery of the human past, which doesn’t become less complex or less perplexing as we back up to take a wider view.
The second set of misunderstandings I want to address at the outset involves a polemical definition of “distant reading” that Franco Moretti advanced about nineteen years ago. I have embraced the term “distant reading” because it is apt, and because I am wary of the academic tendency to simultaneously disavow and appropriate the past by rebranding it. (“Everyone knows that distant reading was naïve, but I have invented critical distant reading, which is quite another matter!”) Endless rebranding is tiresome. However, it needs to be said that the way to evaluate distant reading, in 2019, is to look at the results recently produced by a growing community of scholars—not to stage a debate with a speculative rationale for this project that Franco Moretti put forward in the year 2000.
Moretti was not the first scholar to propose exploring the literary past with social-scientific methods and digital texts. Similar projects, inflected by corpus linguistics, sociology, and book history, were already under way in the 1980s and 1990s. The project accelerated dramatically at the beginning of this century, fueled by a set of social and conceptual innovations that could support large-scale research (digital libraries, for instance, and machine learning). But few of those factors were visible to most literary scholars in the year 2000. Instead, distant reading was initially understood as an extension of the canon-expanding recovery projects of the 1990s. This gave the enterprise a moral claim on scholars’ attention. If you didn’t do distant reading, you were presumably ignoring the cries of thousands of volumes forgotten in “the slaughterhouse of literature.”
Nineteen years later, the project of large-scale literary history is still often called distant reading because the phrase is vivid and appropriate. But the project has outgrown the polemics that originally accompanied its name. For instance, Moretti’s emphasis on the moral urgency of recovery prompted many skeptics to reply that digital libraries themselves still exclude many volumes that are either lost or simply not digitized. No collection, however large, can save every work from the slaughterhouse. This is true. It is also not an objection to contemporary practices of distant reading, which usually work with explicitly limited samples. The point of distant reading is not to recover a complete archive of all published works but to understand the contrast between samples drawn from different periods or social contexts.
In this and many other ways, distant readers and their critics are often simply talking past each other. Quantitative approaches to literary history have been quite productive, but the results they have produced are not the results predicted by their most notorious manifestoes. The differences between the canon and the slaughterhouse, for instance, turn out not to be enormous. Prominent and obscure writers are often moving in roughly the same direction. But in expanding the scope of their analyses, distant readers have stumbled onto long historical arcs that change what we thought we knew about both groups of writers, canonical and obscure.
It is time for this conversation to refocus. Distant readers need new manifestoes that provoke critics to respond to what they have actually done—which might be even more interesting than what they had promised to do two decades ago. This book refocuses the conversation in one of several possible ways, by shifting emphasis from sheer archival comprehensiveness to the sweep of long timelines.
The methods I will be describing do, of course, have limits. It would be a mistake to push numbers into every corner of literary study merely because they are new and fun. Critics who want to sensitively describe the merits of a single work usually have no need for statistics. Enthusiasm for computers and glossy pictures has sometimes led observers to overstate how much can be added to our understanding of a single book by, say, network graphs detailing the connections between its characters. Computational analysis of text is more flexible than it used to be, but it is still quite crude compared to human reading; it helps mainly with questions where the evidence is simply too big to fit in a single reader’s memory. This is why quantitative methods have contributed especially to our understanding of long timelines.
On the other hand, a book about literary history cannot spend all its time thirty thousand feet above the ground. Literature grips readers through individual characters and resonant details; literary history needs to do the same thing. This is especially true for a history of modern literature. As chapters 1 and 3 explain, concrete specificity has become steadily more important to poetry and fiction across the last three hundred years and constitutes at present the main stylistic difference separating literary genres from nonfiction. A history of modern literature that confined itself to sweeping generalization would fail to convey a crucial dimension of its subject. So, while taking a very wide view of history, this book does also plunge into case studies of individual authors and close readings of selected passages.
The rhetorical and aesthetic strains created by this juxtaposition of scales pose the real challenge for distant reading. There may be no conflict, in principle, between quantitative reasoning and humanistic interpretation. But it remains true that literary scholarship aims at an aesthetic standard more exacting than the one prevailing in science. Can distant readers write quantitative literary history that is nevertheless detailed enough, streamlined enough, and lively enough to interest a wide range of readers? If we can’t, then no argument will save us: what we are doing may be important, but it will belong in the social sciences. I hope to show that numbers can also be at home in the humanities. But I cannot prove that in advance. I can only aspire to demonstrate it by writing a book that uses statistical models to tell a suspenseful story of broad human interest.