Theater as Data

Front matter

Computational Journeys into Theater Research

Miguel Escobar Varela

COMPUTATIONAL JOURNEYS INTO THEATER RESEARCH

Part 2: Guided Tours 4 Words as Data 5 Relationships as Data 6 Motion as Data

Part 3: Ensuring the Journeys Continue 8 The Imperative of Open and Sustainable Data viii e Contents

Appendix A: Data Biographies 189 Appendix B: Technical Glossary 193 References 195 Index 219

Digital materials related to this title can be found on the Fulcrum platform via the following citable URL: https://doi.org/10.3998/mpub.11667458

I am on my way to the theater, and I can’t stop generating data. When I use public transport, consult a map on my phone, and order a coffee at the theater foyer, Iam leaving digital breadcrumbs of my activities behind me. The data points that have resulted from these activities will be aggregated with those generated by thousands of other people. Collectively, these datasets will shape the experience of riding buses, consulting maps, and ordering coffee for myself and others. From a technological perspective, it is doubtlessly exciting that data can be applied to such diverse aspects of life. But there are many complex ethical and political corollaries to this intense collection and use of data. Data is not only useful for optimizing bus routes and coffee pricing, it can also be used to increase social inequality (O’Neil 2016, 94) and to sway elections (Morgan 2018). We live in a world that is made for data and from data, as data shapes us in more ways than one. As Ribes and Jackson (2013) note, “we have entered into a symbiotic relationship with data—remaking our material, technological, geographical, organizational, and social worlds into the kind of environments in which data can flourish” (52).

I finish my coffee, switch off my phone, and enter the performance space. It seems like I have entered a data-free sanctuary and left the world of data and its ominous potentials behind. After all, the creative decisions that shape performances are most likely not driven by data. While data-driven content is becoming increasingly common in films and television (Suri and Singh 2018), and data plays a role in the marketing and selection of performances, I can’t find any documented instance where the decisions about what happens on stage were shaped by data, not even in commercial productions. There are many performances which are about data, such as Of All of the People in All of the World by Stan’s Café (2008),

2 œe THEATER AS DATA where grains of rice are used to represent population statistics. However, creative decisions (lighting, casting, modes of interacting with the audience) are not decided by data. This is certainly not the case in the intimate, experimental productions and cultural performances to which the bulk of scholarly attention is devoted to. Likewise, most theater scholarship is focused on specific, thoroughly contextualized performance events. Theater and performance scholars deploy interpretive methods to study the performances that interest us. We don’t usually make use of data—in the information science sense—in order to seek answers to our questions or to orient our discussions. That is, until recently.

With the advent of the digital humanities (henceforth DH), and the pioneering work of several researchers, using data is becoming more common in theater scholarship. This does not mean surrendering our methods to the nefarious potentials of data, but combining data and critical insight to offer fresh perspectives. For example, Holledge et al. (2016) have used a comprehensive dataset of A Doll’s House performances to offer new insights into the global spread of Ibsen’s masterpiece. Using network analysis, Trilcke et al. (2015) have identified previously unreported patterns in the interaction of characters in German drama across several centuries. Schéch (2017) has identified clear genre differences in the words used by classical French playwrights. Caplan (2017) has used network data to show that Yiddish theater performers are far more influential in popular entertainment that they are usually credited for. Wiesner and Stalnaker (2015) were able to estimate conceptual metaphors in dance from context-free movement patterns. And Miller (2017) has used Broadway data to show that the performance runs of musicals have consistently increased in length over time.

These are just some examples of work carried out at the intersection of theater and DH, an area with a long history. In an overview of this intersection, Spence (2013) outlines many projects that were decades in the making. In her overview of this area, Leonhardt (2014) describes projects that fit into seven categories: library and archival projects, projects on individual playwrights or actors, projects on theater architecture, projects on dance, larger network projects, applied DH in theater research and education/experiential DH projects in performance studies, and projects useful for or related to theater and performance research. Caplan (2015) distinguishes between four types of projects: digital archives and editions, digital theatrical environments, digital visualizations, and digital databases. And Bay-Cheng (2017) classifies existing projects according to three major types of research activities: collection, analysis, and dissemination.

Regardless of how one slices this interdisciplinary cake, one thing is certain: there is a great variety of projects carried out at the crossroads of theater and DH. This variety will feature in this book, but my main objective is not to chronicle this growing area. Rather, I focus on a more specific question: what does it mean to research theater in terms of data? Iam interested primarily in studying theater as an event rather than as dramatic literature (an area that has been extensively studied elsewhere in DH). I believe, with Peggy Phelan (1993), that theater “constitutes itself through disappearance” (146). As performances blink in and out of existence, they leave traces behind. These include program booklets, play schedules, cast lists, critical reviews, and videos which can be transformed into digital data. Thus, my project is also aligned to what Manovich (2020) calls cultural analytics. While DH is a more expansive term that includes everything from the digitization of cultural heritage to the development of metadata standards, cultural analytics tends to be more narrowly focused on the computational analysis of large cultural datasets. This might well include the analysis of textual materials and historical data, but cultural analytics as practiced by Manovich and his lab since 2005 gives primacy to the analysis of contemporary, visual culture at a scale of millions of objects. In my projects, I also consider visual and other nontextual dimensions of theater, but some of my sources are historical rather than contemporary and my datasets are never as extensive as those used by Manovich.

What Is Data?

This book takes a narrow definition of data as “potential information.” This comes from information science, a field that assumes data is something that needs to be automatically processed to be useful (Pomerantz 2015). Processed data becomes information. Applied information becomes knowledge, and reasoned knowledge becomes wisdom. This hierarchy was famously inspired by T. S. Eliot’s play The Rock (1934):

Where is the wisdom we have lost in knowledge?

Where is the knowledge we have lost in information? (7)

This quote is often found in information science books. When theater scholars turn to other fields for definitions, we might find some comfort in being sent back to look at a theater play. A longer definition of data, as understood in this book, is potential information in digital form that needs to be processed by a machine for it to become meaningful. But the important question, as Christine Borgman (2015) notes, is not “what are data?” but “when are data?” (17). The same digital object could be treated as information or as data. Take, for example, a theater script. If one close reads a theater play, one is not treating it as data, even if the text is stored in a digital format. But if one uses software to calculate the relative frequencies of a word in this text, then the same object is treated as data. The examples of data in this book are digital texts, videos, motion capture, geographical coordinates, and timestamps. In all cases, software is used to analyze and visualize this data. For example, the videos are not viewed by a human observer, but treated as sequences of two-dimensional images made of pixels. The distribution of pixels is used to make inferences about the ways actors move around the stage.

This narrow definition of data will help bring focus to the arguments in this book, but I will not ignore the many problems it poses. For example, when data is processed, it is often described as “flowing” from one form into another. However, as Jonathan Bollen (2017) notes, “if the metaphor applies at all, datasets have the fluidity of plastic; they flow when prodded, pushed, or pressed, when the heat is turned up and when placed under stress” (619). Data is never fully free of interpretation. As Lisa Gitelman (2013) writes: “Data are familiarly ‘collected,’ ‘entered,’ ‘compiled,’ ‘stored,’ ‘processed,’ ‘mined,’ and ‘interpreted’ [...] less obvious are the ways in which the final term in this sequence—interpretation—haunts its predecessors” (3). The history of pre-digital data further illustrates some of the problems associated with data. Daniel Rosenberg (2013) argues that data has often referred to aspects of an argument taken for granted for the sake of discussion or analysis. “The semantic function of data,” he tells us, has always been “specifically rhetorical” (8, original emphasis). In one of its earlier recorded uses, Euclid differentiates the given quantities of data from the questia of the quantities sought (Rosenberg 2013, 19). These views evolved during the seventeenth century, when the term became more common:

[In] philosophy and natural philosophy, just as in mathematics and theology, the term “data” functioned to identify that category of facts and principles that were, by agreement, beyond argument. In different contexts, such agreement might be based on a concept of self-evident truth, as in the case of biblical data, or on simple argumentative convenience as in the case of algebra, given X=3, and so forth. The term “data” itself implied no ontological claim. In mathematics, theology, and every other realm in which the term was used, “data” was something given by the conventions of argument. Whether these conventions were factual, counter-factual, or arbitrary had no bearing on the status of givens as data. (Rosenberg 2013, 20)

The prophets of our own current techno-scientific era—scientists, managers, and entrepreneurs—often invite us to treat data as something that is beyond argument. A dangerous corollary of this attitude is the belief that data is always true or accurate. It is often assumed that data is the product of a semi-automatic, or at least systematic process. For example, data can be collected by scientific instruments or sensors. But for the sensors to work in a particular way, a series of assumptions about the phenomenon of interest needs to be made: “data need to be imagined as data to exist and function as such, and the imagination of data entails an interpretive base” (Gitelman 2013, 3 original emphasis). Or, as Borgman (2015) puts it “conceptualizing something as data is itself a scholarly act” (xviii). Data artist Jer Thorp suggests that data is never inseparable from a system where collection, computation, and representation inform each other. The following excerpt from his blog vividly captures the impact of choice on such a system:

Whenever you look at data—as a spreadsheet or database view or a visualization, you are looking at an artifact of such a system. What this diagram doesn’t capture is the immense branching of choice that happens at each step along the way. As you make each decision—to omit a row of data, or to implement a particular database structure or to use a specific color palette you are treading down a path through this wild, tall grass of possibility. It will be tempting to look back and see your trail as the only one that you could have taken, but in reality a slightly divergent you who’d made slightly divergent choices might have ended up somewhere altogether different. (Thorp 2017, n.p.)

In performance studies, Schechner (2013) famously distinguished between “as/is” performance, where certain events are generally thought to be performances whereas others (from political events to sporting competitions) could be thought of as performances (32). Following Gitelman, Borgman and Thorp, I argue that data never is—no natural category of data exists without preconditions. But many things can be considered as data, including theater performances. Perhaps all forms of theater

6 eœ THEATER AS DATA research require us to have an abstract model of what performance is, but conventional scholarly writing allows us to refine and revise definitions as we go along, through the conceptual process that Kwa (2011) describes as a “changing semantic network.” In contrast, data forces us to make definitions explicit. For example, let’s imagine a project that wants to show the average number of actors per performance throughout several decades for a given place. For this analysis, we will need an explicit definition of actor in order to produce the required dataset.

This definition doesn’t need to encompass all the essential qualities of an actor, but it must be narrow and implementable. In other words, it should consist of an actionable rule that enables the systematic creation of a dataset. For example, a research team might define actors, for the purpose of a project, as the people listed as cast members in program booklets. In some cases, this definition will be relatively uncontroversial. But other types of productions will be fraught with more intense disagreement. What about a forum theater performance where the audience members become spect-actors? Should each audience member be included as an actor in the dataset? Or only the resident company actors? Or should a separate category be created for spect-actors? All of these options are problematic, but building the dataset will force its makers to take a stance and decide on definitions of actor. It will be important to make choices that will enable the best inferences to be drawn from the data. For example, if we are interested in analyzing the careers of professional actors, we can leave the spect-actors out. Our data will not be a fully accurate model of reality, but it will be sufficient to find verifiable patterns, through the process that Craig and Greatley-Hirsch (2017) call “principled generalization” (7).

It is, however, also possible to use data in a way that takes this definitional problem head on. Instead of deciding at the outset whether spect-actors should be considered as actors, the goal of a project might be to frame this as a question. A data visualization might be used to explore the consequences of taking spect-actors as actors, showing how different solutions to this definitional problem will yield a completely different number of actors. An interactive implementation of this project might enable users to reach their own conclusions on the appropriate definition of actors. Perhaps the data points in the visualization can be linked to videos, where users can then see the way in which specific spect-actors interact with resident company actors. In this example, data is not used to answer a question, but to reframe the question of what it means to be an actor.

This example shows that there are two ways of working with data. One brackets off assumptions to find patterns, while the other uses data to problematize these assumptions. The former offers new answers, the latter poses the question in fresh ways. I call the first approach data-driven and the second, data-assisted. These two terms form the conceptual backbone of the present book.

Data-Driven and Data-Assisted Methodologies

Data-driven methodologies use computers to reason under formal constraints. Knowledge is conceived as a rational project, where logic and formal representations are used in the pursuit of replicable conclusions that disprove previous ideas. In contrast, data-assisted frameworks use computers to “imagine in different ways” (Harrell 2013, 79). Data opens speculative and subjective avenues for interpretation, which contribute to, but don’t necessarily displace, previous interpretations. Data-driven methodologies aim at producing consistent explanations that are “hard to vary,” as physicist David Detusch (2009) puts it. In turn, data-assisted methodologies produce a multiplicity of explanations.

In data-driven methodologies, we use data to answer questions. We create a formal representation of a question and automate a sequence of procedures to provide an answer. The criteria for evaluation are defined beforehand, and the answer is measured against these criteria. Researchers working in this mode know that framing a question in a narrow way leaves out many aspects of the object of interest, but they believe that there is something to be gained from this reduction. For example, they are able to identify features across a different scale (thousands of performances rather than a single instance). This also enables the possibility of countering our intuitions. This way of working with data brings evenness of attention, creating “an opportunity to be surprised: to back something other than the sentimental favorite and to reverse consensus views” (Craig and Greatley-Hirsch 2017, 8). Using data in this way, we keep our intuitions in check, and independently verify deeply held beliefs. While the answers aim to provide the best way to characterize a phenomenon, they are not final. They can be revised and disproved by further research. They also include probabilistic estimates of their accuracy in order to provide “measured uncertainty” (Craig and Greatley- Hirsch 2017, 3). These estimates can be revised by other researchers as more data becomes available.

In data-assisted methodologies, we use data to transform our view of a problem. The purpose of framing a theatrical event as data is not to offer a clear answer but to augment our capacity to think about such event. Data becomes deformance, to use the portmanteau of deformance and performance suggested by Lisa Samuels and Jerome McGann (1999) to characterize interpretive transformations in DH. Data, in other words, provides a good defamiliarization strategy. Data-assisted methodologies work well with theatrical problems that might be fundamentally unquantifiable. But in trying to quantify them, as in the spect-actors example above, we find new ways to address interesting aspects of these problems. According to Kramnick (2018), the objective of critical method in the humanities is not to give definite answers but to “keep the problem open, poke around its edges, ask whether it has been framed in the right way, resituate the conversation” (n.p.). This is also what data-assisted methodologies do.

The main distinction between these methodologies is the different criteria they require for evaluating conclusions as useful and valid. The gold standard of data-driven research is replicable, incremental knowledge. To assess data-driven claims, we should ask how likely the results are to be true and whether other independent sources of evidence corroborate these conclusions. In contrast, data-assisted frameworks ask to be judged for their generative capacity, their potential to trigger new questions and bring forth new perspectives. While in data-driven methodologies many different procedures and parameters should yield the same answer, in data-assisted frameworks the same approach should yield multiple answers.

Data-assisted methodologies are closer to a constructivist view of knowledge, where every piece of data and every conclusion is observer-dependent. Data-driven methodologies are closer to a realist epistemology. They seek to make intersubjectively verifiable claims, and conceive empirical research as a difficult, but worthy goal. Data-driven approaches consider the advantages of thinking in terms of aggregates, and data-assisted approaches demand close attention to individual instances. Table I summarizes the comparison made thus far between both approaches.

My term data-assisted is a nod to the DH community, where the term “computer-assisted criticism” is common (Siemens 2002; Rockwell and Sinclair 2016). Some of the distinctions between data-driven and data-assisted methods have parallels in the epistemological disagreements within the DH community. But this is not a book about DH as a field. It is a book about what it means to use data to study theater. My starting point, though, is that epistemological discussions arise in any field that deals with data and that a book on data must be a book about methodology. It is disingenuous to think that we can describe data without talking about epistemology.

Thus, an epistemological vein runs through this book, as this is a meditation on how we come to know things about theater and the roles that data might play in this endeavor. For this reason, I have so far been characterizing data-driven and data-assisted approaches as methodologies rather than methods. A methodology is a framework for estimating the pertinence of given methods and for articulating criteria for evaluating results as useful, appropriate, or correct. A method is a protocol, a series of steps. The same method could be used within data-driven and data-assisted approaches. Take, for example, word frequencies. Within a data-driven perspective, one could use word frequencies to identify the most likely author of a dramatic text of unknown authorship. But word frequencies could also be used to radically transform the experience of reading a text. As Stephen Ramsay (2011) suggests, the font size of each word can be altered to represent the relative frequency of that word in the entire text. This data-assisted example doesn’t aim to provide a specific answer to a clear question, but to enable alternative readings via a computational defamiliarization strategy. Both the authorship attribution study and the defamiliarizing change in font sizes might use the same statistical methods for estimating word frequencies.

Statistics are at the heart of many data projects, regardless of whether they are conceived within a data-driven or a data-assisted framework. The presence of statistics does not in itself determine whether a project is

Table 1. Comparison of data-assisted and data-driven methodologies

Data-driven Data-assisted Reasoning under constraints (bracketing Problematizing assumptions assumptions) Answers that are hard to vary Multiplicity of views Productive reductionism Deformance (productive distortion) Replicability and falsifiability Adding to a history of interpretation Estimating the likelihood that an answer Resituating a question is correct Measured uncertainty Fundamentally unquantifiable propositions Realism Constructivism Aggregates Individual cases

Disambiguation Exploration of ambiguity aligned with either methodology. Statistics are not always enlisted for the pursuit of objectivity, as they can also be used as defamiliarization strategies. There is, however, one key difference: while descriptive statistics are used in both approaches, inferential or explanatory statistics belong only to data-driven research. The different roles of statistics are described in more detail in chapter 2.

Visualizations are also central to many data projects, but the way in which they are used varies within both methodologies. Data-driven visualizations rely extensively on statistical graphs, using conventions borrowed from the scientific community. Data-assisted visualizations require the invention of new graphical conventions. Data-assisted visualizations aim to resituate conversations, and to communicate how ideological biases shape the data. Visualizations used in computational theater research can be static or interactive, regardless of which methodology is used. But interactivity is particularly useful for data-assisted visualizations that aim to provide context to each data point, as in the example above, where each spect-actors data point is linked to a video-recording. The different roles of visualizations are described in chapter 3.

When distinguishing between data-driven and data-assisted methodologies, it doesn’t matter whether a project appears to rely extensively on statistics or not, or the extent to which it uses visualizations. The size of a dataset doesn’t matter either: both methodologies can be applied to very small and very large datasets. What matters is why statistics and visualizations are used, and how they support the argument being made. The important thing, in other words, is which criteria for evaluation are enlisted in the communication ofa research project.

There are many differences between data-driven and data-assisted methodologies, but they can also be used concurrently within the same research project, and both can contribute to a more comprehensive understanding of theater history and practice. As noted earlier, computational research exists in a spectrum from realism to constructivism. Data-driven research tends to be closer to realism and data-assisted to constructivism. Some researchers occupy radical positions at the opposite ends of this spectrum, affirming either that “we should strive for a science of culture” or that “data is always observer-dependent and socially constructed.” But it is possible to carry out research from a more nuanced position in the spectrum. I believe, with the proponents of critical realism, that one can be a realist about some things and a constructivist about others. Critical realism is a metatheoretical perspective from the philosophy of science that is becoming increasingly common in the study of the social world. While critical realism is not often invoked in the context of DH, I believe that it is consistent with the nuanced ways in which many researchers speak about their work. For example, many proponents of data-driven methods are quick to recognize that there is a certain subjectivity to their projects. The choice of a research topic, for instance, can’t ever be fully derived from objective first principles and always betrays a situated way of looking at the world. Likewise, many proponents of what I call data-assisted research note that the data used must conform to intersubjectively verifiable standards.

In contemporary humanities departments, calling someone a “positivist” is most certainly an insult. I have witnessed many data researchers defending themselves against accusations of positivistic allegiances by merely stating that no, their projects are not positivistic, but without offering further elaboration. I would like to suggest it is more useful to strive for a fine-grained vocabulary, such as the one that critical realism provides. Critical realism is given more attention in chapter 1, but its presuppositions run through the rest of the chapters in this book. For some projects, data-driven and data-assisted perspectives might coexist. But my goal is not to suggest that every research project must combine data-driven and data-assisted perspectives. Rather, my goal is to encourage any data project to aim towards critical realism. For a data-driven project, this means taking history and definitions seriously. For a data-assisted project, it means aiming to show how playful defamiliarizations and multiple perspectives can contribute to a fuller understanding of a complex, shared reality.

Some projects will be better suited for a data-driven perspective, and others for a data-assisted one. In my view, data-driven research is only useful for research questions that fulfill four conditions:

1. There is low ambiguity in the definitions of key concepts. Take, for example, the analysis of scene structure in theater. In some dramatic traditions, it is possible to establish consensus on what constitutes a scene. This is the case in Ibsen’s plays. But it would be harder to reach consensus in what constitutes scene boundaries in the work of other playwrights.

2. The area of interest can be thought of as discrete features rather than continuous processes. It is possible to analyze methodically the number of productions in which Andre Gregory has been involved as a director. One could create a visualization of his creative output through time. But analyzing the aesthetics of one of his plays would require interpretive work more suited to data-assisted analysis, or other methods.

3. The data is arithmetically malleable. For example, one could measure the distance that two touring theater companies covered ina year. One could add these measurements and conclude that one company traveled twice as far as the second one. But one could not say that one company is twice as original as another, even in cases where the critical consensus is that one company is indeed more original.

4. Data can be combined into an aggregate. For certain research projects, it might make sense to take the entire choreographic oeuvre of Akram Khan as a unit. But it might not make sense to consider migrant theater as a unit. Perhaps what makes migrant theater interesting as a site of study is the diversity of individual approaches.

These four conditions are related, but they are not equivalent to each other. One could imagine a research area where there is high ambiguity, but researchers can still identify discrete, arithmetically malleable features. That is, many researchers identify discrete features but they don’t agree on what they are. Let’s take, for example, the plays attributed to Shakespeare. There is some disagreement on which plays he wrote, but plays are still discrete features that are arithmetically malleable. Different researchers can offer competing numbers of plays written by Shakespeare. The analysis of areas such as this, which only meet some of the four criteria, are better served by data-assisted methodologies, provided that discrete features can still be identified (otherwise there is nothing to be turned into data). Cases where none of the criteria can be identified are better served by other types of methodologies that don’t use data, such as ethnography or practice-based research. This information is summarized in table 2. Iam sure that some people will disagree with my examples above. The-

Table 2. Criteria for different types of research

Criteria Data-driven Data-assisted Other approaches Low ambiguity Yes Maybe No Discrete features Yes Yes No Arithmetic malleability Yes Maybe No

Aggregability Yes Maybe No ater and performance studies are, after all, built around sophisticated disagreement (Strine, Long, and Hopkins 1990). Perhaps there are researchers who consider the question of Shakespeare authorship settled, scholars of Ibsen who could point me to disputed scene boundaries in his plays, or dance scholars who believe that Akram Khan’s work is too diverse to be considered as an aggregate. My proposed criteria are not meant to be taken as inherent properties of a research area, but as articulations of a researcher’s perspective. Data-driven research is useful for someone who assumes that the four criteria are met for a given research project. Researchers who believe that such assumption would be misleading should better use a data-assisted methodology. Assuming that the four criteria are met doesn’t mean sweeping problems under the rug, and low ambiguity doesn’t mean no ambiguity. A researcher can still acknowledge disagreement in her field, while choosing to pursue a data-driven project. This acknowledgment can be communicated together with the results of the research. Or the consequences of this disagreement could be further analyzed through a data-assisted methodology.

In addition to the criteria described above, every data project requires good sources of data. The availability of data will always be connected to the existence of specific records, and this will limit the kinds of questions we can ask. The kinds of data available also signal the existence of specific histories of representation, and might raise many critical issues that should be addressed when using the data. Good data is also thick data. We need to know how and why the data was collected, and how it was processed. I use the term thick data following Tricia Wang (2016), who extends the anthropological concept of thick description to data. Thick data should also include thorough descriptions of our doubts and misgivings in the collection of data, a chronicle of the decisions made and a description of the roads not taken. Data biographies are useful for communicating these aspects of data and whenever possible, they should accompany a research project (more on this on chapter 8).