Computational Parallax as Humanistic Inquiry

Humanistic Parallax

European astronomers in the seventeenth century faced a situation familiar to computational humanists: their instruments and methods were embroiled in controversies over disciplinary territory and status at the university, definitions of their objects of study, and reliability of techniques for reconciling mathematical observations with received historical and theoretical understandings of these objects (Westman, 230–34). In particular, the astronomers debated parallax, a theorized effect that would allow for modeling the cosmos using the calculation of the distance between celestial objects based on measurements related to observer positions on earth. To make effective use of parallax, the fractious natural philosophers would have needed to agree on the overall dimensions of their objects of study, establish an accurate underlying model of their relationships in time, and adhere to stable taxonomies of those objects. For textual scholars, these astronomical disagreements might echo current literary and historical questions about the ideal collection of texts or corpus to study, which features determine a meaningful relationship between texts, and the consistency or even validity of genre categories. This chapter does not propose a solution to these long-standing debates but rather a method for expanding humanistic interpretation of text using computation, in spite of the unsettled nature of the textual environment in which we operate.

This method takes its inspiration from the fusion of astronomy and literary studies that was frequently practiced in seventeenth-century Europe. The provisional solution to the astronomers’ apparent impasse was not mathematical but humanistic, in the European Renaissance sense of the term: the examination of particulars of textual variants in search of common contexts and stability of relationships across documents (Mueller, 83).1 Rather than wait for agreement on dimensions, models, and categories of objects (which would only arrive centuries later), the unsigned, jocose “Dialogue of Cecco di Ronchitti” (1604), often attributed to Galileo Galilei, shifts the goal of parallax away from measurement toward aggregating relative attributes of the objects of study using these early humanistic methods. In the dialogue, two mathematically inclined farmers experience difference instead of calculating it by viewing trees in a forest from different perspectives. Their geometric, not measured, experiments lead to the declaration: “Parallax just means a difference of viewpoint [deferientia de guardamento]” (Galileo, 48). To be more precise, parallax is the accumulation of attributes of an object when seen in multiple contexts generated by an observer’s changed position. Importantly, it means keeping our (computational) eye trained on one object while we change the ways in which we observe it or the backdrops against which it is compared.

Computational humanities research frequently makes reference to parallax, but with inconsistent definitions that often foreclose a study of ambiguity, polyphony, and critical imagination that are bedrock elements of much current humanistic study. Adapting the metaphor from the farmer-astronomers in 1604, calculating the forest often obscures our ability to experience a tree. I argue for framing computational experiments that rely on parallax for interpretation with the recovered late-Italian Renaissance articulation of the roles of the observer, the objects of study, and contexts. I am intentional about referring to this as computational rather than digital parallax. Admittedly, parallax can be created digitally through graphical display tools. This can be through a background that scrolls more slowly than a foreground in digital storytelling, or through interfaces that overlay or juxtapose partial views of an object’s colocation in a collection (Whitelaw, 36–37). For my purposes, computation refers to repeated, complex calculations of relationships between humanistic objects of study. Those calculations, made for multiple models that reflect changes in corpora, parameters, or scale, artificially create a sense of space in which the objects exist. The analyst, then, experiences and interprets that space.

Adopting the visual, geometric framework of parallax that emphasizes the analyst’s change in perspectives unites existing threads of discourse in digital humanities (DH). As such, computational humanistic parallax responds to Johanna Drucker’s related suggestion that “parallax and difference” in computational methodologies could emphasize variability, irregularity, and complexity as opposed to representations of humanistic data that otherwise risk a “self-evident or self-identical presentation of knowledge” in diagrammatic or numeric results (“Humanistic Theory,” 91). Computational humanistic parallax offers a valuable heuristic for interpreting expression by units of analysis that exist within and across collections of texts: varying backdrops (different corpora), multiple perspectives (parameters of observation), and attention to scale (levels of addressability). I will describe each of these features after offering an overview of the appearance of parallax in other digital humanities research.

Importantly, my synthesis comes from a white, multilingual settler perspective that is privileged and not all-encompassing. Right to left (RTL) and multilingual DH are actively resisting the limitations of previous computational humanistic approaches and building alternative designs that include rather than exclude perspectives for analysis.2 Moreover, the scholarship that I present reflects the languages that I can navigate, and given the information architectures that create filters in the name of expediency, it is quite likely that even the examples here are but a limited portion of what exists in those critical traditions. This limitation invites expansion from colleagues within and outside the Anglophone and Romance language traditions of text analysis. Does my argument nonetheless move away from a universalizing, normalizing approach to data? Yes, I think so. Does it give agency to more voices in the documents? Here, too, I am optimistic that it creates possibilities to explore connections foreclosed by post hoc categories imposed by majority-organized archives and information practices.

Survey of the Field

In DH, the concept of parallax has been deployed already as a way to add perspectives and layers to computational humanistic analysis, with a particular emphasis on change over time and the use of extratextual data for adding depth to resulting models. Johanna Drucker, using graphics credited to Xárene Eskandar, initially presented graphical design interventions for representing parallax of witnesses and participants in narrative events (“Humanities Approaches,” 36). Stephen Nichols later proposed parallax as layering and juxtaposition of a digitized library of manuscript variants, which is quite aligned with the method for using parallax adopted by Galileo’s farmers. Mark Sample has presented parallax as a hermeneutic for scalable reading by placing one text in the contexts of different corpora in order to ask questions about the multiple meanings that a work offers to readers over time, here too maintaining the object of study while changing parameters for analysis and interpretation. Ted Underwood most recently explores parallax as perspectival modeling of genre labels over time through the use of machine learning (Distant Horizons and “Machine Learning”). Nonetheless, most of these publications include a section of defense that this computational work is humanistic, writing in anticipation of a presumed fractious audience.

By recovering the Renaissance humanistic definition of parallax adopted by Galileo’s farmers, current humanists need not agree on the definitive model of the largest dataset; we can instead direct our computational energy toward the ways in which an individual object participates (or not) in larger patterns. Such a framework also overcomes a frequent challenge of quantitative approaches to interpreting remixed and remixable ecosystems of expression: a struggle to expand beyond sociological or historical conclusions. The recovered definition reasserts the scholar’s role in creating the contextual conditions for observing relationships and the multiple, simultaneous ways that text carries or expresses meaning. This multiplicity has been identified using different descriptors across the humanities: from “massively addressable” text (Witmore) to polyrhythmic performances (Andrews) to images and spaces designed to be viewed and experienced from multiple perspectives (Ruecker and Roberts-Smith). Rather than computation aiming for contextualization or definition within the largest dataset possible, computational humanistic parallax aims for the largest possible variation of contexts and features to analyze. The definition that I propose emphasizes the scholar’s active role in collating these perspectives on their objects of study through multiple observations. The scholar is also participant, experiencing the results of the change as a way to interpret the object, in line with, but not limited to, Sample’s changing corpora. This approach expands Drucker and Eskandar’s initial conceptualization of parallax beyond the content of the humanistic object to the process of its analysis.

Computational parallax as a process thus declares its conditionality and constructedness at the outset in ways that align with traditional humanistic inquiry. As Drucker reminds us: “The parallax views that arise in the interstices of fragmentary evidence are what give humanistic thought its purchase on the real, even with full acknowledgment that knowing is always a process of codependencies” (“Humanistic Theory,” 92). From the perspectives offered by computational humanistic parallax, we have learned something that can add to our understanding of our text(s). For example, a sentiment analysis model might assign a numeric score to a passage to represent positive or negative valence, based on the sum of the scores of the words it contains. Changing passage sizes and textual environments, we can interpret the passage based on a multiplicity of results revealed by computation, without affixing a measurement to a passage as though it could only have one way of making meaning. This is one way a computational tool can work in service of humanistic concerns.

This focus on the relationship of one document to the rest of the corpus sets computational humanistic parallax apart from the related practices of exploratory data analysis (EDA). Both are iterative and rely on an attitude of experimentation rather than the application of a theory to produce multiple visualizations and analyses of a dataset. Yet, EDA’s goal is to summarize the data as a whole through multiple methods of summarizing, aggregating, and visualizing variables, their categories, or their relationships before determining which statistical model(s) will best describe the objects of study. On the other hand, computational humanistic parallax makes no presumption about the possibility of creating such a model and instead draws attention to the relationship of the part to the whole as it is defined at the time, without claims to definitiveness of the corpus of texts.

Since the status of an optimized model is debatable, much like the astronomical theories that provoke Galileo’s farmers, computational humanistic parallax offers a mechanism for interpreting the individual pieces of that system. It is a mechanism that both assists in reproducibility and makes evident when computational results cannot be compared because all the conditions of the original observations cannot be re-created. This separates model creation from interpretation, following Drucker’s advocacy away from output that risks being mistaken as a representation instead of an analytical result that requires explanation. Our interpretations might still vary, depending on the other knowledge that we bring to these patterns and representations. This framing also moves us away from reductive criticisms that computation of humanistic study can only show aboutness and its quantities, influence and its strengths, and the stability or change of categorical labels over time (Da, 605). For example, if I change the corpus but keep the parameters and scale the same, then I can interrogate why the results might be different rather than argue that a model is wrong. Elsewhere in this volume, Mark Algee-Hewitt advocates for a similar use of unexpected quantitative results to inform textual theory (see chapter 2). Building on these examples, I want to advance the argument that the priorities of humanistic interpretation can drive computational methods through a closer examination of computational humanistic parallax at work in corpus design, parameter variation, and scales of address.

Backgrounds and Corpus Construction

Unlike Galileo’s farmer-astronomers, who had an established forest for their experiential experiments, current humanists recognize how incomplete their inventory of objects might be. Computational humanistic parallax calls for designing multiple corpora in order to account for both the outsized role of certain texts that obscure our vision and the systemic forces that exclude other texts from analysis. As Lauren Klein has demonstrated, documenting absence requires disrupting the (digital) archive that is often complicit in the process of silencing some voices. Disruption requires interpretation, not declaration: a computational practice whose aim is not classification or testing conformity to a category but a process that creates analytical space.

Instead of assembling the largest environment for analysis, computational parallax invites assembling multiple possible scenarios for that environment. No matter how large the corpus, how big the big data, or how small the passage, many computational approaches such as clustering will only show the best fit based on a strongest signal determined by the features of the other texts being analyzed. To avoid this outcome, for his parallax reading, Mark Sample places one of Theodore Roethke’s poems in biographical, chronological, and spoken contexts alongside keyword searching in Google Ngram Viewer and the Corpus of Historical American English. For my project, rather than optimizing one model to document Galileo’s overall position among texts in the sixteenth and seventeenth centuries, I have compared his texts against the backdrops of pre-Reformation fiction, prose by Florentine authors, Italian dialogues, and Aristotelian mathematical treatises. Consider two visual examples of the overall similarity between one of Galileo’s texts to seventy-three other documents circulating during his lifetime (Figure 4.1a) and to only those twenty-eight documents published in the seventeenth century (Figure 4.1b). The focus here is on the contingent nature of that similarity when the same analysis is performed on both corpora; a more detailed analysis follows in the next sections.

Many attributes have been collapsed into two dimensions in these visualizations (more on this in the next section). Thinking geometrically, like the astronomer-farmers, I want to draw our attention to the change in relative location (the relative similarity) of Galileo’s text (bold) when the corpus changes. Removing the seventeenth-century documents allows us to see similarities with the earlier texts that would have escaped our attention otherwise. This immediately draws my attention to what meaning might exist in Galileo’s text that originates in the expressions of the century prior in the apparently similar treatise on painting (the point labeled “Della Pittura”). We also cannot help but notice how distant and different the translation of an Arabic treatise on geometry is in Figure 4.1b (the point labeled “Superficie”). These examples are drawn from text analysis, but humanists interpret meaning through recontextualization of a variety of objects of study. As Hannah Ringler argues in chapter 1 in this volume, by seeing the computational tool as a “third eye” in our methodology, we can ask humanistic questions rather than answer quantitative ones with computation.

Perspectives and Parameters

Against these backdrops, the critic can then select from a number of instruments or settings on those instruments to accumulate different perspectives (collapsed into just two opaque dimensions of the x- and y-axes in Figure 4.1). In the example above, each perspective could address a different kind of abstraction of the texts: frequencies, bags of words without order, parts of speech, punctuated interruptions, and metadata. The perspectives show what is present or absent, how much is present in what proportions, along with what other attributes, which layers of significations and in what rhythms, and from what creators or in what containers. Andrew Piper’s work in enumerations embodies much of this parallax spirit by creating different kinds of models to explore questions about the English, German, and French texts under consideration. The computational question is not simply if there are similarities, but under what conditions are similarities observed? For instance, his analysis of plot combines vector space models built from words in context, type-token ratios of quantities of unique and total terms, and network representations of characters named in close context with one another (42–66). In his introduction, Piper is direct about modeling being “contingent world-making that recognizes the situatedness of the critic, aka the model creator, without it being entirely subjective” (12). Some of the tools embed iteration and multiplicity of model building into their design, like topic modeling, which outputs a result optimized for statistical likelihood. Yet, often the inner workings of such output are opaque, such that the attributes that contribute to it are lost to the scholar seeking information about meaning-making.

Here we return to the question of the axes in Figure 4.1. That graph represents the reduction into two-dimensional space of 100 dimensions of information about most frequent words calculated as distance measurements for seventy-five texts. The graph is declarative, but without spaces for asking how it came to be and what that means for the texts it represents. Figure 4.2 provides a simplified example of how parallax can offer such transparency. The ideal output need not be visual, but it does make analysis more experiential, much like the astronomer-farmers walking through their forest. The x and y axes in Figure 4.2 could represent any perspective for analysis: types and tokens, distance from the object of interest in stylometric tests, likelihood of the presence of a topic, frequency of a character’s name or a character pair, or even the reduction to two dimensions of a matrix of quantitative attributes accumulated over iterations of analysis. Here, to keep things within the limits of this chapter, they represent the frequency of “e” (and) and “che” (that/which), two of the most frequent words in Italian texts during Galileo’s period (also two of the reduced dimensions in Figure 4.1). Points are shaped by author: Galileo (triangles) and his favorite poet from a century prior (circles) along with a contemporary poet he loved to hate (squares). To anticipate the question of scale in the last section, the full text measurement is represented by a large shape surrounded by small shapes that represent each section of the respective texts.

It is even easier to see how the sections of Galileo’s work participate differently in the overall relationships in the corpus when the relationships are animated instead of static.3 As the attributes change, we experience the resonance (or not) of the texts with one another across analytical perspectives. The proximity or the spaces made evident by the visualization invite exploration. Similarly, Lisa Rhody’s work on topic modeling figurative language highlights the effectiveness of exploring this interpretative space between the details of a text and the contextual understanding offered by a model. Importantly, I seek ways to identify the place of parts of expression within the whole; Rhody investigates how poetic language resists the model; and Algee-Hewitt pushes us to consider the moments “when the results of our analysis go awry,” which Piper calls the “vulnerability” of texts (159–62). Fragments, resistance, and vulnerability could be extended easily to other forms of expression beyond the texts. Currently, the computational “third eye” can see these relationships, but it only reports an abstraction back to us.

Scale and Addressability

The massive addressability of humanist objects of study immediately challenges the stability implied by describing a genre or period from a collection of feature frequencies that represent only one scale of address. This final aspect of parallax asks us to consider what pieces of the text will be studied: the work in its entirety or segments? Declaring the ability of the object to simultaneously participate in different patterns connects parallax to polyrhythmic performance and analysis. These methods emphasize that aspects of the same expression can expand through time differently. Comparing those permutations through time to others (rather than the object’s persistence through time) allows us to explore what Drucker called “interstices” (“Humanistic Theory,” 92) and Rhody described as “rich deposits of hermeneutic possibility” in the space between what we study and its contexts.

For example, because the richness of literary analysis rests in an acceptance and exploration of the multiple simultaneous associations between texts of varying sizes, the literary scholar can take advantage of the computational ability to create conditions to observe these relationships. I find that using computational parallax for stylometric examination of most frequent words in a corpus of textual sections (the smaller points in Figure 4.2) creates space for a more nuanced interpretation than just which passages are most similar to Galileo’s style. For shorter chapters in his treatise (approximately 500 words), the process reveals rhythmic similarities with other textual segments that have the same rhetorical goals (lengthy description). Yet, similar longer passages (approximately 1,000 words) share narrative devices with its cluster of documents (such as the episodic zigzag from Galileo’s favorite epic poet). Full-text comparisons align the entire work with century-old prose more often than with his contemporaries. We are drawn to these similarities by what John Burrows called “tiny strokes” (268), not explicit sampling or marked, specialist vocabulary that keyword search or topic modeling would identify. Here, they point toward conventions of description, digression, and debate, not traditional genres.

Creating a workflow based on generating parallax and expressing results of varying analytical perspectives thus reduces the possibility of reconfirming interpretations shaped more by the conditions of scholarship than by the expressions of the texts. At the conclusion of her critique of text mining and gender, Laura Mandell gestures toward the value of parallax particularly regarding how some digital humanities research is structured, which risks reification and reconfirmation of categories and definitions that arise post hoc from critics rather than organically from the texts in their multiple contexts and conversations. Mandell, drawing on Donna Haraway, underlines the inordinate power of an observer for defining reality with one statistical result when computation should afford us opportunities for immersion in complexity (Mandell, 17). In chapter 6 in this volume, Katherine McDonough directs us to one solution for breaking out of the imposed or engrained categories of map attributes by using machine learning to look beyond geolocated points to patterns that change how a map can be analyzed, both saving time and possibly stepping away from some of the embedded violence and injustice that a map often represents.

Implications

Using a spatial framework like parallax for computation necessarily relies on understanding text as multidimensional and multicontextual. Interpretations rest not on reduction to a single quantification but aggregation of simultaneous measurements for comparison, contextualization, and evaluation. Expanded research will assess the extent to which this approach holds value for the study of image, sound, and the body. Next steps in computation for this definition of parallax would develop an interface to make more of the objects’ dimensions available for experience and subsequent interpretation. The ARTFL project has anticipated some of these needs with the experimental reading environment for The Intertextual Hub, which allows for toggling between results of different reading tools for documents and corpora as a way to navigate their collection of texts. This need not be supported by a consortium like ARTFL, since the code exists in a shareable Jupyter notebook currently, but it will require collaboration across the spectra of programming and humanistic specialties to move to a more user-friendly phase with the ease and functionality of something like a Voyant Tools widget. Much like Galileo’s farmers walking through an orchard rather than measuring it, the work to be done involves prioritizing experience over calculation. By using the recovered humanistic definition of parallax to guide computation, computation itself is no longer the goal, replaced by identifying as many possible perspectives on the forest as we care to explore in order to better understand the trees within them.

Notes

1. 1. See chapter 15 in this volume.

2. 2. RTL scholars have offered workshops at the Digital Humanities Summer Institutes in Victoria since 2018 and had been using #Right2Left, #RTL20, #RTL21, etc., to connect RTL speakers and DH researchers. The list multilingual-dh@lists.stanford.edu and the Github repository for Multilingual Digital Humanities (https://github.com/multilingual-dh) offer ways to connect with that community.

3. 3. See animated graphic at https://www.bowdoin.edu/~chall/ParallaxAnimated.mov (also available as a gif).

Bibliography

1. Andrews, Richard. Polyrhythmicity in Language, Music and Society: Complex Time Relations in the Arts, Humanities and Social Sciences. Singapore: Springer, 2021.
2. ARTFL. “The Intertextual Hub. Search and Navigation across Digital Collections.” Accessed July 30, 2021. https://intertextual-hub.org/.
3. Burrows, John. “‘Delta’: A Measure of Stylistic Difference and a Guide to Likely Authorship.” Literary and Linguistic Computing 17, no. 3 (2002): 267–87.
4. Da, Nan Z. “The Computational Case against Computational Literary Studies.” Critical Inquiry 45 (Spring 2019): 601–39.
5. Drucker, Johanna. “Humanities Approaches to Graphical Display.” DHQ: Digital Humanities Quarterly 5, no. 1 (2011). http://www.digitalhumanities.org/dhq/vol/5/1/000091/000091.html.
6. Drucker, Johanna. “Humanistic Theory and Digital Scholarship.” In Debates in the Digital Humanities, edited by Matthew K. Gold, 85–95. Minneapolis: University of Minnesota Press, 2012. https://dhdebates.gc.cuny.edu/read/untitled-88c11800-9446-469b-a3be-3fdb36bfbd1e/section/0b495250-97af-4046-91ff-98b6ea9f83c0#ch06.
7. Galilei, Galileo. “The Dialogue of Cecco di Ronchitti.” In Galileo against the Philosophers, edited and translated by Stillman Drake, 33–53. Los Angeles: Zeitlin and Ver Brugge, 1976.
8. Klein, Lauren. An Archive of Taste: Race and Eating in the Early United States. Minneapolis: University of Minnesota Press, 2020.
9. Mandell, Laura. “Gender and Cultural Analytics: Finding or Making Stereotypes?” In Debates in the Digital Humanities 2019, edited by Matthew K. Gold and Lauren F. Klein, 3–26. Minneapolis: University of Minnesota Press, 2019. https://dhdebates.gc.cuny.edu/read/untitled-f2acf72c-a469-49d8-be35-67f9ac1e3a60/section/5d9c1b63-7b60-42dd-8cda-bde837f638f4#node-a34ec8b32594cb722950f75611f4038a278e7464.
10. Mueller, Paul R. “Textual Criticism and Early Modern Natural Philosophy: The Case of Marin Mersenne (1588–1648).” In The Word and the World. Biblical Exegesis and Early Modern Science, edited by Kevin Killeen and Peter J. Forshaw, 78–90. New York: Palgrave Macmillan, 2007.
11. Nichols, Stephen G. “The Anxiety of Irrelevance: Digital Humanities and Contemporary Critical Theory.” Poetica 45, no. 1/2 (2013): 1–17.
12. Piper, Andrew. enumerations. Chicago: University of Chicago Press, 2018.
13. Rhody, Lisa. “Topic Modeling and Figurative Language.” Journal of Digital Humanities 2, no. 1 (2012). http://journalofdigitalhumanities.org/2-1/topic-modeling-and-figurative-language-by-lisa-m-rhody/.
14. Ruecker, Stan, and Jennifer Roberts-Smith. “Experience Design for the Humanities: Activating Multiple Interpretations.” In Making Things and Drawing Boundaries, edited by Jentery Sayers, 259–70. Minneapolis: University of Minnesota Press, 2017. https://doi.org/10.5749/j.ctt1pwt6wq.34.
15. Sample, Mark. “A Parallax Reading of Roethke’s ‘My Papa’s Waltz.’” samplereality.org, May 31, 2017. https://www.samplereality.com/author/admin/page/2/.
16. Underwood, Ted. Distant Horizons: Digital Evidence and Literary Change. Chicago: University of Chicago Press, 2019.
17. Underwood, Ted. “Machine Learning and Human Perspective.” PMLA 135, no. 1 (2020): 92–109.
18. Westman, Robert. The Copernican Question. Berkeley: University of California Press, 2011.
19. Whitelaw, Mitchell. “Generous Interfaces for Digital Cultural Collections.” DHQ: Digital Humanities Quarterly 9, no. 1 (2015). http://www.digitalhumanities.org/dhq/vol/9/1/000205/000205.html.
20. Witmore, Michael. “Text: A Massively Addressable Object.” In Debates in the Digital Humanities, edited by Matthew K. Gold. Minneapolis: University of Minnesota Press, 2012. https://dhdebates.gc.cuny.edu/read/untitled-88c11800-9446-469b-a3be-3fdb36bfbd1e/section/402e7e9a-359b-4b11-8386-a1b48e40425a.