48

Do Digital Humanists Need to Understand Algorithms?

Benjamin M. Schmidt

Algorithms and Transforms

Ian Bogost recently published an essay1 arguing that fetishizing algorithms can pollute our ability to accurately describe the world we live in. “Concepts like ‘algorithm,’” he writes, “have become sloppy shorthands, slang terms for the act of mistaking multipart complex systems for simple, singular ones” (Bogost). Even critics of computational culture succumb to the temptation to describe algorithms as though they operate with a single incontrovertible beauty, he argues; this leaves them with a “distorted, theological view of computational action” that ignores human agency.

As one of the few sites in the humanities where algorithms are created and deployed, the digital humanities are ideally positioned to help humanists better understand the operations of algorithms rather than blindly venerate or condemn them. But too often, we deliberately occlude understanding and meaning in favor of an instrumental approach that simply treats algorithms as tools whose efficacy can be judged intuitively. The underlying complexity of computers makes some degree of ignorance unavoidable. Past a certain point, humanists certainly do not need to understand the algorithms that produce results they use; given the complexity of modern software, it is unlikely that they could.

But although there are elements to software we can safely ignore, some basic standards of understanding remain necessary to practicing humanities data analysis as a scholarly activity and not merely a technical one. While some algorithms are indeed byzantine procedures without much coherence or purpose, others are laden with assumptions that we are perfectly well equipped to understand. What an algorithm does is distinct from, and more important to understand, than how it does it. I want to argue here that a fully realized field of humanities data analysis can do better than to test the validity of algorithms from the outside; instead, it will explore the implications of the assumptions underlying the processes described in software. Put simply: digital humanists do not need to understand algorithms at all. They do need, however, to understand the transformations that algorithms attempt to bring about. If we do so, our practice will be more effective and more likely to be truly original.

The core of this argument lies in a distinction between algorithms and transformations. An algorithm is a set of precisely specifiable steps that produce an output. “Algorithms” are central objects of study in computer science; the primary intellectual questions about an algorithm involve the resources necessary for those steps to run (particularly in terms of time and memory). “Transformations,” on the other hand, are the reconfigurations that an algorithm might effect. The term is less strongly linked to computer science: its strongest disciplinary ties are to mathematics (for example, in geometry, to describe the operations that can be taken on a shape) and linguistics (where it forms the heart of Noam Chomsky’s theory of “transformational grammar”).

Computationally, algorithms create transformations. Intellectually, however, people design algorithms in order to automatically perform a given transformation. That is to say: a transformation expresses a coherent goal that can be understood independently of the algorithm that produces it. Perhaps the simplest example is the transformation of sorting. “Sortedness” is a general property that any person can understand independently of the operations that produce it. The uses that one can make of alphabetical sorting in humanities research—such as producing a concordance to a text or arranging an index of names—are independent of the particular algorithm used to sort. There are, in fact, a multitude of particular algorithms that enable computers to sort a list. Certain canonical sorting algorithms, such as quicksort, are fundamental to the pedagogy in computer science. (The canonical collection and explanation of sorting algorithms is the first half of Knuth’s canonical computer science text.) It would be ludicrous to suggest humanists need to understand an algorithm like quicksort to use a sorted list. But we do need to understand sortedness itself in order to make use of the distinctive properties of a sorted list.

The alternative to understanding the meaning of transformations is to use algorithms instrumentally; to hope, for example, that an algorithm like Latent Dirichlet Allocation will approximate existing objects like “topics,” “discourses,” or “themes” and explore the fissures where it fails to do so. (See, for example, Rhody; Goldstone and Underwood; Schmidt, “Words Alone.”) This instrumental approach to software, however, promises us little in the way of understanding; in hoping that algorithms will approximate existing meanings, it in many ways precludes them from creating new ones. The signal criticism of large-scale textual analysis by traditional humanists is that it tells scholars nothing they did not know before. This critique is frequently misguided; but it does touch on a frustrating failure, which is that distant reading as commonly practiced frequently fails to offer any new ways of understanding texts.

Far more interesting, if less immediately useful, will be to marry large-scale analysis to what Stephen Ramsay calls “algorithmic criticism”: the process of using algorithmic transformations as ways to open texts for new readings (Ramsay). This is true even when, as in some of the algorithms Ramsay describes, the transformation is inherently meaningless. But transformations that embody a purpose themselves can help us to create new versions of text that offer fresh or useful perspectives. Seeking out and describing how those transformations function is a type of work we can do more to recognize and promote.

The Fourier Transform and Literary Time

A debate between Annie Swafford and Matt Jockers over Jockers’s “Syuzhet” package2 for exploring the shape of plots through sentiment analysis offers a useful case study of how further exploring a transformation’s purpose can enrich our vocabulary for describing texts. Although Swafford’s initial critique raised several issues with the package, the bulk of her continuing conversation with Jockers centered on the appropriateness of his use of a low-pass filter from signal processing as a “smoothing function.” Jockers argued it provided an excellent way to “filter out the extremes in the sentiment trajectories.” Swafford, on the other hand, argued that it was often dominated by “ringing artifacts” which, in practice, means the curves produced place almost all their emphasis “at the lowest point only and consider rises or falls on either side irrelevant” (Jockers, “Revealing Sentiment”; Swafford “Problems”; Swafford, “Why Syuzhet Doesn’t Work”).

The Swafford and Jockers debate hinged over not just an algorithm, but a concretely defined transformation. The discrete Fourier transform undergirds the low-pass filters that Jockers uses to analyze plot. The thought that the Fourier transform might make sense as a formation for plot is an intriguing one; it is also, as Swafford argues, quite likely wrong. The ringing artifacts that Swafford describes are effects of a larger issue: the basic understanding of time embodied in the transformation itself.

The purpose of the Fourier transform is to represent cyclical events as frequencies by breaking complex signals into their component parts. Some of the most basic elements of human experience—most notably, light and sound—physically exist as repeating waves. The Fourier transform offers an easy way to describe these infinitely long waves as a short series of frequencies, constantly repeating. The pure musical note “A,” for example, is a constant pulsation at 440 cycles per second; as actually produced by a clarinet, it has (among other components) a large number of regular “overtones,” less powerful component notes that occur at a higher frequency and enrich the sound beyond a simple tone. A filter like the one Jockers uses strips away these regularities; it is typically used in processes like MP3 compression to strip out notes too high for the human ear to hear. When applied even more aggressively to such a clarinet tone, it would remove the higher frequencies, preserving the note “A” but attenuating the distinctive tone of the instrument.3

The idea that plots might be represented in the frequency domain is fascinating, but makes some highly questionable assumptions. Perhaps the most striking assumption is that plots, like sound or light, are composed of endlessly repeating signals. A low-pass filter like the one Jockers employs ignores any elements that seem to be regularly repeating in the text and instead focuses on the longest-term motions; those that take place over periods of time greater than a quarter or a third the length of the text. The process is analogous to predicting the continuing sound of the clarinet based on a sound clip of the note “A” just 1/440th of a second long, a single beat of the base frequency. This, remarkably, is feasible for the musical note, but only because the tone repeats endlessly. The default smoothing in the Syuzhet package assumes that books do the same; among other things, this means the smoothed versions assume the start of every book has an emotional valence that continues the trajectory of its final sentence. (I have explained this at slightly greater length in Schmidt, “Commodius Vici.”)

For some plots, including Jockers’s primary example, Portrait of the Artist as a Young Man, this assumption is not noticeably false. But for other plots, it causes great problems. Figure 48.1 shows the plot of Portrait and four other novels, with text taken from Project Gutenberg. William Dean Howell’s The Rise of Silas Lapham is a story of ruination; Ragged Dick, by Horatio Alger, is the archetypal “Rags to Riches” novel of the nineteenth century; Madame Bovary is a classically tragic tale of decline. Three different smoothing functions are shown: a weighted moving average, among the simplest possible functions; a loess moving average, which is one of the most basic and least assumption-laden algorithms used in exploratory data analysis; and the low-pass filter included with Syuzhet.4

The problems with the Fourier transform here are obvious. A periodic function forces Madame Bovary to be “as well off” after her death as before her infidelity. The less assumption-laden methods, on the other hand, allow her fate to collapse at the end and for Ragged Dick’s trajectory to move upward instead of ending on the downslope. Andrew Piper suggests5 that it may be quite difficult to answer the question, “How do we know when a curve is ‘wrong’?” (Piper, “Validation”). But in this case, the wrongness is actually quite apparent; only the attempt to close the circle can justify the downturn in Ragged Dick’s fate at the end of the novel.

What sort of evidence is this? By Jockers’s account,6 the Bovary example is simply a negative “validation” of the method, by which I believe he means a sort of empirical falsification of the claim that this is the best method in all cases (Jockers, “Requiem”). Swafford’s posts imply similarly that case-by-case validation and falsification are the gold standard. In her words, the package (and perhaps the digital humanities as a whole) need “more peer review and rigorous testing—designed to confirm or refute hypotheses” (Swafford, “Continuing”).

Seen in these terms, the algorithm is a process whose operations are fundamentally opaque; we can poke or prod to see if it matches our hopes, but we can never truly know it. But when the algorithm is a means of realizing a meaningful transformation, as in the case of the Fourier transform, we can do better than this kind of quality assurance testing; we can interpretively know in advance where a transformation will fail. I did not choose Madame Bovary at random to see if it looked good enough; instead, the implications of the smoothing method made it obvious that the tragedy, in general, was a type of novel that this conception of sentiment that Syuzhet’s smoothing could not comprehend. I will admit, with some trepidation, that I have never actually read either Madame Bovary or Ragged Dick; but each is the archetype of a plot wholly incompatible with low-pass filter smoothing. Any other novel that ends in death and despair or extraordinary good fortune would fail in the same way.

These problems carry through to Jockers’s set of fundamental plots: all begin and end at exactly the same sentiment. But the obvious problems with this assumption were not noted in the first two months of the package’s existence (which surely included far more intensive scrutiny than any peer-review process might have). One particularly interesting reason that these failings were not immediately obvious is that line charts, like Figure 48.1, do not fully embody the assumptions of the Fourier transform. The statistical graphics we use to represent results can themselves be thought of as meaningful transformations into a new domain of analysis. And in this case, the geometries and coordinate systems we use to chart plots are themselves emblazoned with a particular model. Such line charts assume that time is linear and infinite. In general, this is far and away the easiest and most accurate way to represent time on paper. It is not, though, true to the frequency domain that the Fourier transform takes for granted. If the Fourier transform is the right way to look at plots, we should be plotting in polar coordinates, which wrap around to their beginning. I have replotted the same data in Figure 48.2, with percentage represented as an angle starting from 12:00 on a clock face and the sentiment defined not by height but by distance from the center.

Here, the assumptions of the Fourier transform are much more clear. For all of the novels here, time forms a closed loop; the ending points distort themselves to line up with the beginning, and vice versa. The other algorithms, on the other hand, allow great gaps: the Madame Bovary arc circles inward as if descending down a drain, and Ragged Dick propels outward into orbit.

These circular plots are more than falsifications. Fully embracing the underlying assumptions of the transform in this way does not only highlight problems with the model; it suggests a new perspective for thinking about plots. This view highlights the gap between the beginning and end as a central feature of the novel; in doing so, it challenges us to think of the time that plots occupy as something other than straightforwardly linear.

This is a conversation worth having, in part because it reminds us to question our other assumptions about plots and time. The infinite time that the Cartesian plot implies is, in some ways, just as false as the radial one. Many smoothing methods (including the one I would like to see used in Syuzhet, loess regression), can easily extrapolate past the beginning and end of the plot. That this is possible shows that they are, in some ways, equally unsuitable for the task at hand. The heart of the distinction between fabula and syuzhet, in fact, is that there is no way to speak about “before the beginning” of a novel, or what words Shakespeare might have written if he had spent a few more hours working past the end of Hamlet. Any model that implies such phrases exist is obviously incorrect.

But even when arguably false, these transformations may yet be productive of new understandings and forms of analysis. While this cyclical return is manifestly inappropriate to the novel, it has significant implications for the study of plot more generally. By asking what sorts of plots of the frequency domain might be useful for, we can abstractly identify whole domains where new applications may be more appropriate.

For example: the ideal form of the three-camera situation comedy is written so that episodes can air in any arbitrary order in syndication. That is to say, along some dimensions they should be cyclical. For sitcom episodes, cyclicality is a useful framework to keep in mind. The cleanness of the fit of sentiment, theme, or other attributes may be an incredibly useful tool both to understand how commercial implications intertwine with authorial independence, or for understanding the transformation of a genre over time. Techniques of signal processing could be invaluable in identifying, for example, when and where networks allow writers to spin out multi-episode plot lines.7

Though the bulk of the Swafford and Jockers conversation centered on the issue of smoothing, many digital humanists seem to have found a second critique Swafford offered far more interesting. She argued that the sentiment analysis algorithms provided by Jockers’s package, most of which were based on dictionaries of words with assigned sentiment scores, produced results that frequently violated “common sense.” While the first issue seems blandly technical, the second offers a platform for digital humanists to talk through how we might better understand the black boxes of algorithms we run. What does it mean for an algorithm to accord to common sense? For it to be useful, does it need to be right 100 percent of the time? 95 percent? 50.1 percent? If the digital humanities are to be a field that appropriates tools created by others, these are precisely the questions it needs to practice answering.

To phrase the question this way, though, is once again to consider the algorithm itself as unknowable. Just as with the Fourier transform, it is better to ask consciously what the transformation of sentiment analysis does. Rather than thinking of the sentiment analysis portion of Syuzhet as a set of word lists to be tested against anonymous human subjects, for example, we should be thinking about the best way to implement the underlying algorithms behind sentiment analysis—logistic regression, perhaps—to distinguish between things other than the binary of “positive” and “negative.” Jockers’s inspiration, Kurt Vonnegut, for example, believed that the central binary of plot was fortune and misfortune, not happiness and sadness; while sentiment analysis provides a useful shortcut, any large-scale platforms might do better to create a classifier that actually distinguishes within that desired binary itself. Andrew Piper’s work on plot structure involves internal comparisons within the novel itself (Piper, “Novel Devotions”). Work like this can help us to better understand plot by placing it into conversation with itself and by finding useful new applications for transformations from other fields.

Doing so means that digital humanists can help to dispel the myths of algorithmic domination that Bogost unpacks, rather than participating in their creation. When historians applied psychoanalysis to historical subjects, we did not suggest they “collaborate” with psychoanalysts and then test their statements against the historical record to see how much they held true; instead, historians themselves worked to deploy concepts that were seen as themselves meaningful. It is good and useful for humanists to be able to push and prod at algorithmic black boxes when the underlying algorithms are inaccessible or overly complex. But when they are reduced to doing so, the first job of digital humanists should be to understand the goals and agendas of the transformations and systems that algorithms serve so that we can be creative users of new ideas, rather than users of tools the purposes of which we decline to know.

Notes

Bibliography