The Signal in the Noise: Hermeneutics and/of Computational Literary Sociology

1 Critique and Postcritique

The tradition of hermeneutics, according to Paul Ricœur, originates between two poles (1970, 26 sq.). One, rooted in scriptural exegesis, restores meaning through interpretation; the other, traceable in the work of Marx, Freud and Nietzsche and called by Ricœur the hermeneutics of suspicion, demystifies in order to uncover hidden meaning. As this hermeneutic has been taken up by literary scholars, texts and the language they are made of have become symptomatic of meanings available for decoding by the critic. In his articulation of the tenets of what is often called ideology critique, Fredric Jameson elaborates on Louis Althusser’s identification of history as an absent cause in order to argue that this cause is »inaccessible to us except in textual form« and that »our approach to it […] necessarily passes through its prior textualization, its narrativization in the political unconscious« (1982, 35). This treatment of texts as symptomatic of an absent historical cause implicitly authorizes readings like Sinykin’s, where texts produced in the context of increasing conglomeration appear as symptoms of these changes. As Sinykin explains it: »I do not mean that authors kept these questions [about profitability, spin-offs, and other pressures related to conglomeration, A. M.] in mind when writing […]. Success depended on recognition by something like a system, so much so that fiction itself, when published by conglomerates, came to display, seen as a whole, a systematic intelligence, a systematic authorship« (2023, 103). The scale shift here, from individual text to historical publishing context, is made possible through something like Jameson’s political unconscious, which allows the »fiction itself« to reveal something about the context of its production. This constitutes a »negative hermeneutic function« insofar as it engages in »demystifying« or »unmasking« a consciousness not immediately visible on the surface of the text (Jameson 1982, 291).

This hermeneutic tradition has had a privileged relationship with the work of literary sociology coming out of US literary studies departments. In fact in James English’s 2010 account, the sociology of literature has suffered as an independent field in the US because Jamesonian approaches to texts have made every literary scholar a sociologist, whose aim, across »postcolonial studies, queer theory, new historicism«, was »to provide an account of literary texts and practices by reference to the social forces of their production, the social meanings of their formal particulars, and the social effects of their circulation and reception« (viii). Jameson himself suggests that for Marxism »literary and cultural analysis is a social science« (1981, 297). Most importantly, what Jameson values in critique as he outlines it is that it »liberates us from the empirical object, […] displacing our attention« from the text toward »its constitution as an object and its relationship to the other objects thus constituted« (ibid., emphasis original). Contrary to the postcritical position that frames ideological critique as a reduction of the text to a repeated symptom of the same hidden cause, Jameson here articulates an expansion of the text to include its political and sociological context. However, desire for liberation from the text as »empirical object« betrays the fact that the text acts as a mere material witness to the larger historical conflict whose elucidation is the scholar’s real aim.

Since around the turn of this century, and particularly as a result of its perceived appropriation by the right in the form of conspiracy theories, critique has increasingly become problematized as scholars seek alternative, non-suspicious relationships to texts. In Eve Kosofsky Sedgwick’s contribution to this genre, she desires a positive hermeneutic – reparative rather than paranoid – in which the text not only reveals previously hidden knowledge (which, in Sedgwick’s telling, is often not very hidden after all), but also provides the reader with something new, like knowledge of how best to »move among [knowledge’s] causes and effects« (2003, 124). Sociologist Bruno Latour similarly provides a model for a positive hermeneutic in which »the critic is not the one who debunks, but the one who assembles«, where humanities scholarship becomes »associated with more, not with less, with multiplication, not subtraction« (2004, 246, 248). This emphasis on quantity – the rhetoric of a critic as someone who »adds reality to matters of fact« rather than »subtract reality« (ibid., 232) – seems to anticipate the introduction of quantitative literary sociology into this debate. The positivist aspect of postcritical approaches often goes hand in hand with attention to a text’s surface, which, in Stephen Best and Sharon Marcus’s explication requires taking the text on its own terms, accepting that it does as it says rather than treating it as a symptom to be decoded or as an opaque container of hidden meanings (2009, 1 sq.).

Both critique and postcritique assert themselves as particularly sociological, but at different epistemological scales: for Jameson, textual critique constitutes a kind of sociology, whereas for Rita Felski, who has promoted a Latourian approach to literary studies, tracing social connections forms the basis for claims about textual meaning (2015, 11). Felski insists that the »connections and mediations« that produce, circulate, and remediate a text must be »tracked down and described« in order to produce »detailed accounts of the actors, groupings, assemblies, and networks« that would provide evidence for the causal connections between the text and history that ideological critique takes for granted (ibid.). The text as such is an active part of a network, which, in concert with various other actors (publishers, editors, readers, books, etc.), creates literary history, rather than presenting as a symptom of that network.

Computational approaches have been identified as a »worthy fellow traveler in the move from depth to surface reading« (Bode 2023, 519; see also Love 2010, 382 sq.). Best and Marcus present computation as a way of avoiding a hermeneutic of suspicion as »computers are weak interpreters but potent describers, anatomizers, taxonomists« (2009, 11 sq., 17). Underwood also sees computational modeling as descriptive and »deeply compatible with an antifoundational approach to interpretation« (2019, 67). »Computer-generated quantitative scholarship«, Felski writes, is »not free of the long shadow of suspicion, but [it] does not orient [itself] toward critique as [its] primary rationale and vindication« (2015, 26). Computational literary sociology, then, is regularly allied with surface reading, description, and non-suspicious reading practices, primarily because computational models are not understood to have the capacity for hermeneutic maneuvering – they take texts literally, word for word.

Postcritical scholars posit that ultimately it is the specificity and (lack of) strength of their theories that sets them apart from those engaged in critique, a distinction often described in terms of scale. Sedgwick, following Silvan Tomkins, uses the language of weak and strong theories (language not meant to be normative) in order to classify the scalability of causal explanations (2003, 133 sq.). Theories that offer repeatable large-scale explanations for global phenomena such as ideology critique are strong theories (to adapt her language to existing allegorical sociological readings: if everything can be understood as manifesting the conflict of the literary field, the conflict of the literary field is everywhere [ibid., 135]). Those that provide small-scale explanations for local phenomena – little more than a »description of the phenomena which [the theory] purports to explain« – are weak. Sedgwick herself argues for a weak reparative reading practice with minimal explanatory power. Felski similarly urges postcritical scholars to take not the birds-eye view of »the social« but rather to construct it from the ground up (she wants not to »soar like an eagle«, but to »trudge along like an ANT«) (2015, 157 sq.).

Computational literary sociology observes texts at a scale inaccessible to human readers (much more similar to an eagle), framing losses of specificity as a cost for generalizability and the »strength« of the theories it produces (Piper 2016, 6). Moretti famously calls this approach »distant reading« where distance is »a specific form of knowledge« (2007, 1, emphasis original), accessible only through an approach that takes texts as empirical instances of larger historical phenomena. What has emerged here is a portrait of a method at odds with one that, limited by its computational objectivity, remains on the surface, reading only texts as they are. In addition to operating on a distant scale, aiming thereby to produce strong and generalizable theories, computational literary sociology also treats texts as empirical objects, data points in the investigation of the distinguishing features of a corpus. In other words, texts become material witnesses in an unfolding historical drama, recalling the symptomatic treatment of the text against which postcritical scholars normally orient themselves. Is it possible that, counter to postcritical scholars’ embrace of computational methods, computers really aren’t postcritical after all? Is the computer a hermeneut?

2 The Statistical Hermeneutic

By taking the very small scale – punctuation marks, words, word groupings – as its basis for claims about the large scale – genres, periods, reception – computational literary sociology blows up the importance of small parts of a text while making descriptions of the social appear manageable. This relationship between the small and large scale in computational literary sociology is revealed in the first instance through statistical modeling. But the belief that taking a representative sample and running a statistical regression – a technique that determines the strength and nature, positive or negative, between a dependent variable and one or several independent variables – can provide insights into human behavior or cultural production is, like the assumptions made by critique or postcritique, historically contingent. It relies on and shapes a particular kind of human subjecthood and stance toward the text that appears objective only insofar as it remains unhistoricized.

When used in computational literary sociology, statistical modeling operates under the twin assumptions that something meaningful about cultural production and reception can be described through quantitative measurement and that statistics elicit meaning from within data that cannot be gleaned by the human mind alone (Nelson 2022, 857 sq.). Statistical methods are not particularly recent, dating in some respects to Galileo, who used them to determine which of a set of measurements of difficult-to-observe phenomena was most probably correct (Smith 2019, 76). When, as Simone Murray (2023) writes, »scientific positivism exerted epistemological force across disciplines«, sociology underwent a process, »whereby the anecdotalism of the Victorian man of letters was being replaced by statistical data newly available to industrialised, increasingly urbanised societies with urgent issues of public health, underemployment and mass education in need of evidence-based policy solutions«. Within this historical context, statistics came to be understood quite differently. Rather than being treated as a way to interpret measurements, statistics became a way to interpret the society that those numbers aimed to measure. This transformation introduced a normative and ideological angle into the observation and analysis of human behavior (and was what made it ideal for its use in eugenics [Smith 2019, 84; see also Gould 1981]).

This transformation required a new belief in what has become a fundamental conceit of statistical models of society, namely that meaningful norms, patterns, or distributions are created by human behavior, and that these are discoverable through statistical modeling. Just as in the European Enlightenment, when humans conveniently became autonomous creatures designed to use reason to discover the world just as the world was becoming rational and law-abiding, the statistical revolution produced an image of society extremely well-matched to the tools recently engaged to measure it.

This statistical hermeneutic has, with the rise of rapid, algorithmic computation of data, begun to shape our quotidian realities, registering within our daily lives in the language of metrics, predictions, and optimization; it has become the defining feature of the way we conceptualize the world in the computational age. Norbert Wiener, writing in 1950, traces this understanding to the discovery of quantum physics (1989, 10 sq.). Where physics used to be concerned with what will always happen, suddenly, physics became concerned with »what will happen with an overwhelming probability« (ibid.). In a discipline that aims to observe and measure reality, »chance has been admitted, not merely as a mathematical tool for physics, but as part of its warp and weft«, altering the way we understand our relationship to the world (ibid.). Kate Crawford sees the origin of a »faith that mathematical formalisms would help us understand humans as a society«, in military-funded research of signal processing and optimization during World War II (2021, 213). »The belief that accurate prediction is fundamentally about reducing the complexity of the world gave rise to an implicit theory of the social«, Crawford writes of this paradigm shift, »find the signal in the noise and make order from disorder« (ibid.). The popular application of statistics in explaining human behavior is now so ubiquitous that it has become banal. Consider, for example, the rise of statistics in the sports industry, where »savvy application of large databases of detailed historical data allows […] analysts to more precisely quantify (and, at the negotiation table, price) the past and potential future impact of a player, lineup, or tactic on outcomes, adjusting for numerous variables, which are multiplying as new tools measure ever more granularly what happens during games« (Burks 2023, 56). Our conceptualization of the world as made meaningful by statistics has brought with it a nearly insatiable desire for optimization made visible in sports and elsewhere with the constant passive capturing of data like keystrokes made, steps taken, and pages turned, as well as the implementation of computational aids toward efficiency and profit maximization.

Cultural and literary production are not immune from these effects. Richard Jean So writes that cultural institutions are using data to »analyze, market and sell novels« (2020, 184). Melanie Walsh and Maria Antoniak suggest that Amazon Studio’s productions are made with an eye toward Goodreads data, citing the filmed adaptations of Philip K. Dick’s The Man in the High Castle and J. R. R. Tolkien’s The Lord of the Rings trilogy (2021, 273 sq.). That Amazon Publishing uses customer data (which it presumably models statistically in order to make understandable at a human scale) in making publishing decisions is, in fact, part of its marketing strategy (Montgomery 2019). We could, therefore, imagine a world in which literary production was primarily controlled not by human selection, but by the same types of predictive models that determine, for example, what sets apart a prize-winning novel. In this scenario, predictive models that determine what distinguishes a prize-winning novel would eventually become tautological: their outputs would become the same as their inputs. The data points would cluster ever more closely around the regression line, and there would be no noise: only a signal being transmitted and received by a machine. The noisy variation – the messiness of human agency – is the problem that statistical models aim to »solve«.

In modern computational literary sociology, statistical patterns are identified through tools that convert the blurriness, uncertainty, and variability of outcomes that occur when measuring human behavior into a clearer, more certain, single abstraction. The p-value, for example, is a predetermined value that works to signal whether or not a model’s output is statistically significant. Like language, this value is arbitrary but conventional and communicates to practitioners the relative strength of the pattern they have identified. By rewriting a textual corpus on the scale of human understanding, this process – homologized by the regression line produced through statistical conventions and traced over a collection of dots – acts as a hermeneutic, indicating the presence of meaning where there was none before. This statistical hermeneutic allows scholars to uncover or produce – the distinction turns out to be precisely the one generating the debate between critique and postcritique – a signal in the noise, for units like genre and period. Underwood writes of this process that »numbers can extract a trend from noisy variation« (2018, 350, emphasis A. M.). For So, data is »a lot of noise« that can be turned into »hard, meaningful signals« (2020, 18).

The repetition of the phrase »signal in the noise« conjures up the idea of deep space exploration – of listening for some sign of intelligence in the seemingly random thrum of the universe. Is there something out there? If a signal were to be received, signifying the presence of intelligence, it would then also become a symbol, requiring interpretation in order to become meaningful. On the one hand, a signal transmits information on its surface: on or off, stop or go, presence or absence. A symbol, on the other, always stands for something. A statistical model’s output functions as a signal that tells us that there is information to be gleaned from the visual noise. But it is also a symbol (Andrew Piper calls it a translation [2018, x]) that asks us to interpret that information and turn it into (human, usually causal) meaning, understandable in (human) language. The frequent use of the word signal centers the (fairly new) assumption that seemingly entropic human behavior always already contains invisible information, but eschews the fact that, as Tarleton Gillespie writes, these signals are »hieroglyphs: shaped by the tool by which they are carved, requiring of priestly interpretation« (2014, 190). In other words, the output of statistical models contains both: a signal that produces order from disorder, and a symbol that prompts renewed hermeneutic attention to fulfill the promise of meaning that the signal extends.

The outputs of computational models, in mapping correlations with diverse relative strengths, provide a (historically contingent) description of the world, hence their perceived affinity with surface reading. Yet, to consider this hermeneutic postcritical misses that statistics is always already a strong theory that creates global, large-scale explanations through the demystification of seemingly random noise. Underwood’s use of the word »extract« demonstrates that computational models test relationships between variables internal to a corpus in order to unearth a signal that was presumably there all along.

Perhaps counterintuitively it is the second hermeneutic maneuver, the human translation – the work of interpreting the output as symbol – that acts as a »black box« in computational literary sociology. Whereas the assumptions made in the collection of data, the variables used, and the code for the models can be made explicit, the verbal narrative that aims to explain what the signal means requires – like critical readings of texts – an interpretive leap. The impulse to generalization, far from having been introduced by computational methods, has been part of the aims of literary studies since at least the advent of new historicism (Piper 2016, 4 sq.). But the interpretive leap required to move between the small and the large – suggesting, for example, that a dialogue in a Shakespeare play dramatizes the conflict of the play’s production at the Globe – is not in fact overcome by the suggestion that the changing frequency of a particular group of words or type of sentence across twentieth century mystery novels tells us something about the desires of readers, the pressures of the market, or the changing structure of the publishing industry. The direction of cause and effect is different here – how do textual markers that have created a signal at scale signify something about the social, rather than how does the social signify within the text – but the mechanism and the interpretive leap is the same. Moretti reveals this unity when he calls distant reading the »completion« of the hermeneutic project: »because the ›abstract objects‹ computation produces – these objects no one experiences directly, but which we all somehow know to take into account – are exactly what hermeneutics wants to raise to the level of consciousness« (2017, 7). We are starting on the opposite bank, but we still need to ford the river. Yes, computational literary sociology operates at scale, but there is a fundamental homology between critique and computation in their shared acts of interpreting a signal that becomes evidence of some meaningful (but hidden) historical cause. The question the statistical model as symbol asks is: how did the signal get into the noise?

3 Reading the Signal in the Noise

Interpretation does not always mean critique, and there are a range of approaches to interpreting the outputs presented by these models and therefore answering the question of how the signal got into the noise. It is precisely because these models are not »objective«, but tell us something about subjective assessments of prestige, literariness, fictionality or whatever else that the question of how these subjective assessments make their way into the language of the texts becomes interesting and important. What mechanism produces this signal?

For Underwood the question isn’t answerable based on a reading of the model’s output alone. The models themselves provide description, but there is no proof »that the inner workings of the model mirror deep structures of the world« (2019, 190, emphasis A. M.). The turning away from depth echoes that of the postcritique where meaning is made through description and connection (»predictive models describe a relationship between social and textual evidence« [ibid., 24, emphasis A. M.]), rather than by unearthing a relationship between texts and the world. Together with Sellers, Underwood creates a model that is able to predict with 79.2 % accuracy whether or not a piece of fiction or poetry published between 1820 and 1919 was reviewed in a prestigious venue. In doing so, they confirm that »diction is meaningfully related to reception« (2016, 328, emphasis A. M.). But this is where the questions begin rather than end: »how can a statistical model be right 79.2 percent of the time? And what was the secret to getting reviewed in this period? We have to«, they write, »content ourselves with a rough answer« (ibid., 330). The answer is rough because it contains too many variables for us to understand in their complexity, so Sellers and Underwood provide an interpretation that points to concrete language as a key indicator of reviewability from among the model’s outputs.

It is clear that the results of this model incite interpretation, even if its makers finally abjure the act itself. Part of what is meaningful to Sellers and Underwood is the discovery that a single model can be used to predict the reviewability of texts across a whole century, which they interpret to mean that tastes change more slowly than scholars previously thought. They also discover that the »model trained on 1845–69 sees works from the 1870s as more likely to be reviewed than the works it was trained on« (Underwood/Sellers 2016, 336 sq.). This provokes the »wild efflorescence of interpretive work« that Nick Seaver (2022, 113), an anthropologist of algorithms, says is prompted by the outputs of computational models:

We didn’t expect to see this, and we don’t want to claim that we understand why it happens. We might speculate, for instance, that standards tend to drift upward because critics and authors respond directly to pressure from reviewers or because they imitate, and slightly exaggerate, the standards already implicit in prominent examples. In that case, synchronic standards would produce diachronic change. But causality could also work the other way: a long-term pattern of diachronic change could itself create synchronic standards if readers in each decade formed their criteria of literary distinction partly by contrasting »the latest thing« to »the embarrassing past.« In fact, causal arrows could point in both of these directions. (Underwood/Sellers 2016, 336 sq.)

Underwood and Sellers ultimately say that »causal processes are hard to trace in detail« and »we [don’t] actually need a causal explanation of this phenomenon« (ibid, 337). Here, the signal has been received, but its significance remains uncertain.

Karina van Dalen-Oskam is similarly reticent in offering causal explanations for her computational findings, although her descriptions of them often highlight illuminating outliers. For example, after she describes the result that statistical models find patterns in texts that allow them to attribute authorship, which is a situation in which we know the signal being produced is meaningful because we can corroborate its production, she acknowledges that »we cannot yet determine exactly how it works, because there are so many different variables involved. For the time being, however, most researchers assume that the differences can partly be traced back to the unique development of the language skills of each individual author« (2021, 21). Even when we know that the signal is an accurate representation of reality, how it got there requires interpretation, because we can’t say definitively why one author writes differently from another, and yet consistently with her past self. Van Dalen-Oskam’s interpretive approach is often to highlight cases that models do not predict correctly. In discussing predictive modeling work done with her student Andreas van Cranenburgh, she writes »it is fascinating to see which titles were being misjudged [by the model]« (ibid., 265), and she suggests that models that predict genre could be productively compared to the genre labels assigned by the marketing division of publishing companies in order to »show where tension arises between marketing strategy and textual form and content« (ibid., 183). Here it is not the signal, but the texts that can’t be correctly ordered by the signal that might produce the most meaning.

Moretti is much more willing to cross the interpretive gap that exists between the signal and the symbol. Unlike Underwood, Moretti usually proposes an a priori interpretation of his findings, namely that through readers’ selections, signals in quantitative data are produced through a sort of evolutionary competition of forms in which they »fight for the limited resources of the market« (2013, 103). »If there is a pattern in the data«, Moretti writes of the signal, »it’s because behind it there is a form which repeats itself over and over again« (2017, 6, emphasis original). Moretti’s evolutionary explanation has been critiqued elsewhere (see for example Prendergast 2005), and Moretti’s explanatory emphasis on readers’ choices self-consciously ignores the possible effects of »publishing, and distribution and their various appendices (reviewing, advertising, etc.)« (2013, 140 sq.).

Yet Moretti’s approach retains the form of a hermeneutics of suspicion, epitomized in his language which places meaning behind the signal, and in which the signal needs to be demystified in order to reveal the conflict of the literary marketplace. Moretti, however, acknowledges that his interpretations are generally just hypotheses because it is unclear that readers, who are supposedly selecting the features they prefer, are even able to see the variables that the models identify (the conceit, remember, is that we cannot see these repeated features without the leverage that computation affords us). Again, the relationship between the signal and the signal’s inscription is a »black box« (this is, in fact, the term Moretti himself uses [2013, 144]). Whereas Underwood suggests that this interpretive gap might be bridged by qualitative scholarship (2019, 139), Moretti suggests that an answer might lie in cognitive science, which could tell us how the many different components of texts are interpreted without our conscious knowledge by the human brain, and therefore how the signal gets into the noise in the first place (2013, 144).

In a study that correlates the setting of novels (future, present, or past) with whether they are bestselling or prize-winning, English interprets the signal first critically and then postcritically to arrive at two different (although not necessarily incompatible) meanings. After describing his results – that around 1980 historical fiction became frequently awarded by prize committees while becoming less frequent on bestseller lists – English states that: »it remains, of course, to explain this transformation« (2016, 411, emphasis original). He first explains it by engaging »Jameson’s causal argument about the cultural effects of neoliberalism since the 1970s« (ibid., 412). However, after butting up against the problem that ideology critique would read in any signal English would have found »the favored tropes of the late capitalist era as symptoms of false historical consciousness« (ibid.), he turns instead to Latourian Actor-Network theory by tracing reviews in prestigious journals and watershed prizes, building a network that ends up centering on the publication of Salman Rushdie’s Midnight’s Children. Here, English suggests that »the critical preference for novels set in the past may not have been imposed by grand, hidden forces but was constructed by a fairly small number of well-placed actors and their concrete interactions at the ground level« (ibid., 415). English’s self-conscious reinterpretation here highlights the necessity of interpreting the model’s statistical description of the world, while also revealing the agnosticism of computational literary sociology in terms of its allegiance to either a critical or postcritical interpretation of that description.

So’s Redlining Culture, which combines qualitative and computational approaches to make claims about race in the corpus of fiction published by Random House in the second half of the twentieth century, is instructive for its treatment of a model’s »noise«. So takes a large-scale view to establish that racial inequality exists in Random House’s publishing lists, describing this corpus »at a cognitive scale well beyond what a single person can observe or read« (2020, 6). The model that So develops is able to predict which books are bestsellers and which are by Black writers 93 percent of the time, which »means that bestselling novels emit a textual signal«, one that »we likely didn’t see before«, »that makes it different from other kinds of novels, novels by black authors in particular« (ibid., 124 sq.). This textual signal (or at least part of it) is adverbs: bestselling novels use them at a higher rate than Black authors. So interprets this accumulation of adverbs in bestselling novels as »a constant deferral of signification« that renders the voice of the American bestseller »an overthinking one« (ibid., 124). His reading here is of what the signal means for the texts, but what does it mean for the production of those texts? How did this adverbial signal enter the noise? Here is So’s interpretation:

Most likely the institutional and textual dynamic exist within a feedback loop. Each day, an editor reads dozens of submissions in search of the next blockbuster or bestselling novel. Our analysis says that this editor, consciously or unconsciously, does this work in part by noticing a set of textual features that are simultaneously very common in previous bestsellers and very uncommon in novels written by black writers (ibid., 125 sq.).

There is something like a political unconscious at work here that has inscribed individual historical action such that it is able to be read traditionally as dramatized in the overthinking voice of the bestseller, as well as newly in the meaning of the signal produced by So’s model.

So then, like van Dalen-Oskam, leverages a particular affordance of the data visualization, namely its ability to identify outliers (the noise), to show that Octavia Butler’s Parable of the Sower is incorrectly identified by the model as both a bestseller and a prizewinner even though it belongs in the other group of texts (novels by Black authors). So, in this case, gets close to the text to perform a reading in which the features (primarily adverbs and lack of racialized language) that the model usually attributes to bestsellers and prize winners rather than novels by Black authors are deployed by the text to subvert the unequal reality of the literary field. Although it may seem like the novel just »adopts the semantic and syntactical style of the bestseller or prizewinner«, So claims that this reading would be »an error in human interpretation« (ibid., 139). Rather, the novel »uses [these] syntactical features – […] otherwise marked by the machine as not distinctive of novels by black authors – precisely to underscore the presence of race in its story frame, albeit in a radically transformed manner« (ibid.). We can’t say how these features – invisible at the human scale, at least consciously – got into the text, but with the help of the machine we can identify them and then interpret them through close reading. So here moves dialectically between the distant and the close, between statistical models, readings of novels, and the archive to suggest interpretations of signals at different scales. He also leverages statistical models to center his attention on the »noise« that becomes a new sort of signal, allowing So to read against, rather than with, the pattern of inequality that he finds.

4 Conclusions

What is computational literary sociology’s stance towards the text? This question is complicated by the fact that the literary text itself often disappears – it is not one of the scales that computational methods usually take on. Instead, they engage in the creation of a signal, epitomized in the example of the data visualization, that connect the scale of the word with the scale of the corpus. The statistical hermeneutic that conditions the inscription of this signal is inherently paranoid, revealing the presence of meaning that was being emitted by human acts of selection all along. The signal/symbol becomes the new object of interpretation, either eliciting critique, in which the symbol’s meaning is demystified in order to reveal the invisible forces that inscribed it, or, prompting surface reading, in which the correlation made visible by the signal is described in a way that draws new connections between previously unattached actors or variables. Despite recentering the text as the origin of the data on which it stakes its claims, computational literary sociology doubles down on the text as a material instantiation of the relationship between human agency and history.

The prevailing assumption that computational literary sociology is unproblematically aligned with postcritical attitudes toward the text misses, due to the fetish of the signal, first, that computational models are themselves never »objective« descriptions of the world but are mediated through a statistical hermeneutic perhaps as ideologically conditioned as critique itself, and second, engages in a type of technological determinism that assumes that these methods can’t be (productively) put to critical uses. Underwood aims to minimize the otherness of computational literary sociology for humanities scholars when he writes that the »twists« of humanistic research can be provided by the theories of critique in which signs prove to mean other than what we thought, or by statistical analysis that shows us that there was really a signal in the noise all along (2019, 166 sq.). When we read the signal in the noise as proving to mean other than what we thought then these two may even at times entirely overlap. For many practitioners of computational literary sociology, the answer to the question »what produces the signal in the noise?« often seems to be something like »the political unconscious«.

This is in no way to disparage either computational literary sociology or critique (or postcritique, for that matter), or to suggest that postcritical scholars who use computational methods are somehow mistaken. It is simply to suggest that the attention that has been paid to the interpretive gap between statistical models and history in critiques of biased/incomplete data collections, errors in digitization, and non-reproducible experiments (a gap that exists in a slightly different form in traditional literary studies as well) would be productively supplemented with closer attention to the nature of the gap itself, as well as to the diversity of complementary methods with which the results of these models can be interpreted. To return to the epigraph in which Wiener identifies mathematics as a metaphor, the sciences are always already ideological in the same way that our reading practices are (even, in my view, those that position themselves against ideological reading) – which ends up being all the more reason to use them in our pursuit of knowledge of ourselves and our cultural production.

Depending on how we read data visualizations – whether on the surface or behind it – we can produce an »assembling« of connections between previously disparate data points, or demystify the symbols that encode previously hidden relationships across long historical narratives. We can describe computational outputs, or we can critique them. But regardless of our reading method what makes these models meaningful is noise. In the hypothetical example of the model that becomes meaningless once its inputs are also selected by a model, not only do the noise and signal become one and the same, but surface reading and critique also become the same. A tracing of a network of actors based on the noise-less input/output of a machine is identical with a narrative uncovering its hidden motives (it has none). The tension, then, between description and critique is in the noise – the ineradicable instability of human selection – and its treatment. Access to and ability to think with this noise in a historical moment that inordinately values statistical optimization is perhaps the most valuable feature of computational approaches to literary sociology today.