How We Read: Close, Hyper, Machine—Part 2December 4, 2013
"How We Read: Close, Hyper, Machine" is republished here from the ADE Bulletin No. 150 (2010) with permission from the author.
Accompanying this article is a portfolio of images from artist Catherine Wagner's 2011 trans/literate series of archival pigment prints, courtesy of the artist and Stephen Wirtz Gallery, San Francisco.
The Importance of Anecdotal Evidence
Faced with these complexities, what is a humanist to do? Obviously, few scholars in the humanities have the time—or the expertise—to backtrack through cited studies and evaluate them for correctness and replicability. In my view, these studies may be suggestive indicators but should be subject to the same kind of careful scrutiny we train our students to use with Web research (reliability of sources, consensus among many different researchers, etc.). Perhaps our most valuable yardstick for evaluating these results, however, is our own experience. We know how we react to intensive Web reading, and we know through repeated interactions with our students how they are likely to read, write, and think as they grapple with print and Web materials. As teachers (and parents), we make daily observations that either confirm or disconfirm what we read in the scientific literature. The scientific research is valuable and should not be ignored, but our experiences are also valuable and can tell us a great deal about the advantages and disadvantages of hyperreading compared with close reading, as well as the long-term effects of engaging in either or both of these reading strategies.
Anecdotal evidence hooked me on this topic five years ago, when I was a Phi Beta Scholar for a year and in that capacity visited many different types of colleges and universities. Everywhere I went, I heard teachers reporting similar stories: “I can’t get my students to read long novels anymore, so I’ve taken to assigning short stories”; “My students won’t read long books, so now I assign chapters and excerpts.” I hypothesized then that a shift in cognitive modes is taking place, from the deep attention characteristic of humanistic inquiry to the hyperattention characteristic of someone scanning Web pages (Hayles, “Hyper and Deep Attention”). I further argued that the shift in cognitive modes is more pronounced the younger the age cohort. Drawing from anecdotal evidence as well as such surveys as the Kaiser Foundation’s Gen M report (Roberts, Foehr, and Rideout), I suggested that the shift toward hyperattention is now noticeable with college students. Since then, the trend has become even more apparent, and the flood of surveys, books, and articles on the topic of distraction is now so pervasive as to be, well, distracting.
For me, the topic is much more than the latest research fad, because it hits me where I live: the college classroom. As a literary scholar, I deeply believe in the importance of writing and reading, so any large-scale change in how young people read and write is bound to capture my attention. In my work on hyperattention (published just when the topic was beginning to appear on the national radar), I argued that deep and hyperattention each have distinctive advantages. Deep attention is essential for coping with complex phenomena such as mathematical theorems, challenging literary works, and complex musical compositions; hyperattention is useful for its flexibility in switching between different information streams, its quick grasp of the gist of material, and its ability to move rapidly among and between different kinds of texts.1 As contemporary environments become more information intensive, it is no surprise that hyperattention (and its associated reading strategy, hyperreading) is growing and that deep attention (and its correlated reading strategy, close reading) is diminishing, particularly among young adults and teens. The problem, as I see it, lies not in hyperattention/hyperreading as such, but rather in the challenges the situation presents for parents and educators to ensure that deep attention and close reading continue to be vibrant components of our reading cultures and interact synergistically with the kind of Web and hyperreading in which our young people are increasingly immersed.
Yet hyper and close reading are not the whole story. I earlier referred to Sosnoski’s definition of hyperreading as “computer-assisted.” More precisely, it is computer-assisted human reading. The formulation alerts us to a third component of contemporary reading practices: human-assisted computer reading, that is, computer algorithms used to analyze patterns in large textual corpora where size makes human reading of the entirety impossible. Machine reading ranges from algorithms for word-frequency counts to more sophisticated programs that find and compare phrases, identify topic clusters, and are capable of learning. Given the scope, pervasiveness, and sophistication of contemporary programs used to parse texts, it seems to me quite reasonable to say that machines can read. One could, of course, restrict “read” to human beings, arguing that reading implies comprehension and that machines calculate but do not comprehend. However, some human readers (beginners, for example) may also read with minimum or no comprehension. Moreover, the line between (human) interpretation and (machine) pattern recognition is a porous boundary, with each interacting with the other. Hypotheses about meaning help shape the design of computer algorithms, and the results of algorithmic analyses refine, extend, and occasionally challenge intuitions about meaning that formed the starting point for algorithmic design. Putting human reading in a leakproof container and isolating machine reading in another makes it difficult to see these interactions and understand their complex synergies. Given these considerations, saying computers cannot read is from my point of view merely species chauvinism.
I agree wholeheartedly with the goal: the question is how, precisely, to accomplish it?
In a field like literary studies, misunderstandings of the efficacy and importance of machine reading are commonplace. Even such a perceptive critic as Culler falls back on caricature when, in writing about close reading, he suggests, “It may be especially important to reflect on the varieties of close reading and even to propose explicit models, in an age where electronic resources make it possible to do literary research without reading at all: find all the instances of the words beg and beggar in novels by two different authors and write up your conclusions” (24). In other words, close reading is the garlic that will ward off the vampire of machine reading. The anxiety here is palpable, nowhere more so than in his final phrase (“write up your conclusions”), which implies that drawing conclusions from machine analysis is a mechanical exercise devoid of creativity, insight, or literary value. Even Guillory, a brilliant theorist and close reader, while acknowledging that machine reading is a useful “prosthesis for the cognitive skill of scanning,” concludes that “the gap in cognitive level between the keyword search and interpretation is for the present immeasurable” (“How” 13). There are two misapprehensions here: that keyword searches exhaust the repertoire of machine reading and that the gap between analysis and interpretation yawns so wide as to form an unbridgeable chasm rather than a dynamic interaction.
Given these misconceptions, explicit recapitulation of the value of machine reading is useful. Although it may be used with a single text and reveal interesting patterns, its more customary use is in analyzing large corpora too vast to be read by a single person. Preeminent in this regard is the work of Franco Moretti, who uses the term “distant reading,” an obvious counterpoise to close reading (Graphs). Careful reading of his work reveals that this construction lumps together human and machine reading; both count as “distant” if the scale is large. I think it is useful to distinguish between human and machine reading because the two situations (one done by a human assisted by machines, the other done by computer algorithms assisted by humans) have different functionalities, limitations, and possibilities. Hyperreading may not be useful for large corpora, and machine algorithms have limited interpretive capabilities.
If we look carefully at Moretti’s methodology, we see how firmly it refutes the misunderstandings referred to above. His algorithmic analysis is usually employed to pose questions. Why are the lifetimes of many different genres limited to about thirty years (Graphs)? Why do British novels in the mid–eighteenth century use many words in a title and then, within a few decades, change so that titles are no more than three or four words long (“Style”)? How to explain changes in narrative conventions such as free indirect discourse when the novel moves from Britain to British colonies (Graphs)? I find Moretti’s work intriguing for the patterns he uncovers, but I am flat out delighted by the ingenious explanations he devises to account for them. So far beyond the mechanical exercises Culler imagines are these explanations that I would not hesitate to call many of them brilliant. When the explanations fail to persuade (as Moretti candidly confesses is sometimes the case even for him), the patterns nevertheless stand revealed as entry points for interpretations advanced by other scholars who find them interesting.
I now turn to explore the interrelations between the components of an expanded repertoire of reading strategies that includes close, hyper, and machine reading. The overlaps between them are as revealing as the differences. Close and hyperreading operate synergistically when hyperreading is used to identify passages or to home in on a few texts of interest, whereupon close reading takes over. As Guillory observed, skimming and scanning here alternate with in-depth reading and interpretation (“How”). Hyperreading overlaps with machine reading in identifying patterns. This might be done in the context of a Google keyword search, for example when one notices that most of the work on a given topic has been done by X, or it might be done when machine analysis confirms a pattern already detected by hyper (or close) reading. Indeed, skimming, scanning, and pattern identification are likely to occur with all three reading strategies; their prevalence in one or another is a matter of scale and emphasis rather than clear-cut boundary.
Since patterns have now entered the discussion, we may wonder what a pattern is. This is not a trivial question, largely because of the various ways in which patterns become manifest. Patterns in large data sets may be so subtle that only sophisticated statistical analysis can reveal them; complex patterns may nevertheless be apprehended quickly and easily when columns of numbers are translated into visual forms, as with fMRI scans. Verbal patterns may be discerned through the close reading of a single textual passage or grasped through hyperreading of an entire text or many texts. An anecdote may be useful in clarifying the nature of pattern. I once took a pottery class, and the instructor asked each participant to make several objects that would constitute a series. The series might, for example, consist of vases with the same shapes but different sizes, or it might be vases of the same size in which the shapes underwent a consistent set of deformations. The example shows that differences are as important as similarities, for they keep a pattern from being merely a series of identical items. I therefore propose the following definition: a pattern consists of regularities that appear through a series of related differences and similarities.
Related to the idea of pattern is the question of meaning.
Related to the idea of pattern is the question of meaning. Since entire books have been written on the subject, I will not attempt to define meaning but merely observe that wherever and however it occurs, meaning is sensitively dependent on context. The same sentence, uttered in two different contexts, may mean something entirely different in one compared with the other. Close reading typically occurs in a mono-local context (that is, with a single text). Here the context is quite rich, including the entire text and other texts connected with it through networks of allusions, citations, and iterative quotations. Hyperreading, by contrast, typically occurs in a multilocal context. Because many textual fragments are juxtaposed, context is truncated, often consisting of a single phrase or sentence, as in a Google search. In machine reading, the context may be limited to a few words or eliminated altogether, as in a word-frequency list. Relatively context-poor, machine reading is enriched by context-rich close reading when close reading provides guidance for the construction of algorithms; Margaret Cohen points to this synergy when she observes that for computer programs to be designed, “the patterns still need to be observed [by close reading]” (59). On the other hand, machine reading may reveal patterns overlooked in close reading, a point Willard McCarty makes in relation to his work on personification in Ovid’s Metamorphosis (53–72). The more the emphasis falls on pattern (as in machine reading), the more likely it is that context must be supplied from outside (by a human interpreter) to connect pattern with meaning; the more the emphasis falls on meaning (as in close reading), the more pattern assumes a subordinate role. In general, the different distributions between pattern, meaning, and context provide a way to think about interrelations between close, hyper, and machine reading.
The larger point is that close, hyper, and machine reading each have distinctive advantages and limitations; nevertheless, they also overlap and can be made to interact synergistically with one another. Maryanne Wolfe reaches a similar conclusion when, at the end of Proust and the Squid, she writes,
We must teach our children to be bitextual or multitextual, able to read and analyze texts flexibly in different ways, with more deliberate instruction at every stage of development on the inferential, demanding aspects of any text. Teaching children to uncover the invisible world that resides in written words needs to be both explicit and part of a dialogue between learner and teacher, if we are to promote the processes that lead to fully formed expert reading in our citizenry. (226)
I agree wholeheartedly with the goal: the question is how, precisely, to accomplish it?
Synergies between Close, Hyper-, and Machine Reading
Starting from a traditional humanistic basis in literature, Alan Liu in the English department at the University of California, Santa Barbara, has been teaching undergraduate and graduate courses that he calls Literature+, which adopt as a pedagogical method the interdisciplinarity facilitated by digital media. He asks students “to choose a literary work and treat it according to one or more of the research paradigms prevalent in other fields of study,” including visualization, storyboarding, simulation, and game design. Starting with close reading, he encourages students to compare it with methodologies in other fields, including the sciences and engineering. He also has constructed a “tool kit” on his Web site that includes links to software packages enabling students with little or no programming experience to create different modes of representation of literary texts, including tools for text analysis, visualization, mapping, and social-network diagramming. The approach is threefold: it offers students traditional literary training; it expands their sense of how they can use digital media to analyze literary texts; and it encourages them to connect literary methodologies with those of other fields they may be entering. It offers close reading not as an unquestioned good but as one methodology among several, with distinctive capabilities and limitations. Moreover, because decisions about how to encode and analyze texts using software programs require precise thinking about priorities, goals, and methodologies, it clarifies the assumptions that undergird close reading by translating them into algorithmic analysis.
An example of how the “Literature+” approach works in practice is the project entitled “Romeo and Juliet: A Facebook Tragedy” (Skura, Nierle, and Gin). Three students working collaboratively adapted Shakespeare’s play to the Facebook model, creating maps of social networks using the Friend Wheel (naturally, the Montagues are all “friends” to each other, and so are the Capulets), filling out profiles for the characters (Romeo is interpreted as a depressive personality who has an obsessive attachment to his love object and who has corresponding preferences for music, films, and other cultural artifacts that express this sensibility), and having a fight break out on the message-board forum using a Group called The Streets of Verona. The Wall feature was used to incorporate dialogue in which characters speak directly to one another, and the Photos section allowed one character to comment on the attributes of another. The masque at which Romeo and Juliet meet became an Event, to which Capulet invited friends in his Friend Wheel. From a pedagogical point of view, the students were encouraged to use software with which they were familiar in unfamiliar ways, thus increasing their awareness of its implications. The exercise also required them to make interpretive judgments about which features of the play were most essential (since not everything could be included) and to be precise about interactions between relationships, events, and characters. Linking traditional literary reading skills with digital encoding and analysis, the “Literature+” approach strengthens the ability to understand complex literature at the same time it encourages students to think reflectively on digital capabilities. Here digital and print literacies mutually reinforce and extend each other.
Lev Manovich’s “Cultural Analytics” is a series of projects that starts from the premise that algorithmic analyses of large data sets (up to several terabytes in size), originally developed for work in the sciences and social sciences, should be applied to cultural objects, including the analysis of real-time data flows. In many academic institutions, high-end computational facilities have programs that invite faculty members and graduate students in the arts and humanities to use them. For example, at the University of California, San Diego, where Manovich teaches, the Supercomputer Center sponsored a summer workshop in 2006, Cyberinfrastructure for the Humanities, Arts, and Social Sciences. At Duke University, where I teach, the Renaissance Computing Institute (RENCI) offers accounts to faculty members and students in the arts and humanities that allow them to use computationally intense analysis. In my experience, researchers at these kinds of facilities are delighted when humanists come to them with projects. Because their mission is to encourage widespread use across and among campuses and to foster collaborations among academic, government, corporate, and community stakeholders, they see humanistic inquiry and artistic creation as missing parts of the picture that enrich the mix. This opens the door to analysis of large cultural data sets such as visual images, media content, and geospatial mapping combined with various historical and cultural overlays.
When it came to digital reading, however, they were accustomed to the fast skimming typical of hyperreading
An example is Manovich’s analysis of Time magazine covers from 1923–89. As Manovich observes, ideal sites for cultural analytics are large data sets that are well structured and include metadata about date, publication venue, and so forth. The visualization tools that he uses allow the Time covers to be analyzed according to subject (for example, portraits versus other types of covers), color gradients, black-and-white gradients, amount of white space, and in other ways. One feature is particularly useful for building bridges between close reading and machine analysis: the visualization tool allows the user both to see large-scale patterns and to zoom in to see a particular cover in detail, thus enabling analyses across multiple scale levels. Other examples include Manovich’s analysis of one million manga pages using the Modrian software, sorted according to gray-scale values; another project analyzes scene lengths and gray scale values in classic black-and-white films. While large-scale data analyses are not new, their applications in the humanities and arts are still in their infancy, making cultural analytics a frontier of knowledge construction.
Of course, not everyone has access to computation-intensive facilities, including most parents and teachers at smaller colleges and universities. A small-scale example that anyone could implement will be helpful. In teaching an honors writing class, I juxtaposed Mary Shelley’s Frankenstein with Shelley Jackson’s Patchwork Girl, an electronic hypertext fiction written in proprietary Storyspace software. Since these were honors students, many of them had already read Frankenstein and were, moreover, practiced in close reading and literary analysis. When it came to digital reading, however, they were accustomed to the scanning and fast skimming typical of hyperreading; they therefore expected that it might take them, oh, half an hour to go through Jackson’s text. They were shocked when I told them a reasonable time to spend with Jackson’s text was about the time it would take them to read Frankenstein, say, ten hours or so. I divided them into teams and assigned a section of Jackson’s text to each team, telling them that I wanted them to discover all the lexias (i.e., blocks of digital text) in their section and warning them that the Storyspace software allows certain lexias to be hidden until others are read. Finally, I asked them to diagram interrelations between lexias, drawing on all three views that the Storyspace software enables.
As a consequence, the students were not only required to read closely but also to analyze the narrative strategies Jackson uses to construct her text. Jackson focuses some of her textual sections on a narrator modeled on the female creature depicted in Frankenstein, when Victor, at the male creature’s request, begins to assemble a female body as a companion to his first creation (Hayles, “Invention”). As Victor works, he begins to think about the two creatures mating and creating a race of such creatures. Stricken with sexual nausea, he tears up the female body while the male creature watches, howling, from the window; throws the pieces into a basket; and rows out onto a lake, where he dumps them. In her text Jackson reassembles and reanimates the female creature, playing with the idea of fragmentation as an inescapable condition not only for her narrator but for all human beings. The idea is reinforced by the visual form of the narrative, which (in the Storyspace map view) is visualized as a series of titled text blocks connected by webs of lines. Juxtaposing this text with Frankenstein encouraged discussions about narrative framing, transitions, strategies, and characterization. By the end the students, who already admired Frankenstein and were enthralled by Mary Shelley’s narrative, were able to see that electronic literature might be comparably complex and would also repay close attention to its strategies, structure, form, rhetoric, and themes. Here already-existing print literacies were enlisted to promote and extend digital literacy.
Reading has always been constituted through complex and diverse practices.
These examples merely scratch the surface of what can be done to create productive interactions between close, hyper, and machine reading. Close and hyperreading are already part of a literary scholar’s tool kit (although hyperreading may not be recognized or valued as such). Many good programs are now available for machine reading, such as Wordle, which creates word clouds to display word-frequency analysis, the advanced version of the Hermetic Word Frequency Counter, which has the ability to count words in multiple files and to count phrases as well as words, and other text-analysis tools available through the TAPoR text-analysis portal. Most of these programs are not difficult to use and provide the basis for wide-ranging experimentation by students and teachers alike. As Manovich says about cultural analytics and Moretti proclaims about distant reading, machine analysis opens the door to new kinds of discoveries that were not possible before and that can surprise and intrigue scholars accustomed to the delights of close reading.
What transformed disciplinary coherence might literary studies embrace? Here is a suggestion: literary studies teaches literacies across a range of media forms, including print and digital, and focuses on interpretation and analysis of patterns, meaning, and context through close, hyper-, and machine reading practices. Reading has always been constituted through complex and diverse practices. Now it is time to rethink what reading is and how it works in the rich mixtures of words and images, sounds and animations, graphics and letters that constitute the environments of twenty-first-century literacies.