Earlier this month, Critical Inquiry hosted an online forum featuring responses to and discussion about Nan Z. Da’s “The Computational Case against Computational Literary Studies.” To accommodate further commentary to Da’s article and to the forum itself, we have created a new page for responses.
- Taylor Arnold (University of Richmond).
- Duncan Buell (University of South Carolina, Columbia).
As a statistician who has worked and published extensively within the fields of digital humanities (DH) and computational linguistics over the past decade, I have been closely following Nan Z. Da’s article “The Computational Case against Computational Literary Studies” and the ensuing conversations in the online forum. It has been repeatedly pointed out that the article contains numerous errors and misunderstandings about statistical inference, Bayesian inference, and mathematical topology. It is not my intention here to restate these same objections. I want to focus instead on an aspect of the work that has gone relatively undiscussed: the larger role to be played by statistics and statisticians within computational DH.
Da correctly points out that computational literary studies, and computational DH more generally, takes a large proportion of its methods, theories, and tools from the field of statistics. And yet, she also notes, scholars have had only limited collaborations with statisticians. It is easy to produce quantitative evidence of this fact. There are a total of zero trained statisticians (having either a Ph.D. or an academic position with the title of statistics) amongst: the 25 members on the editorial board of Cultural Analytics, 11 editors of Digital Humanities Quarterly, 22 members of the editorial board for Digital Scholarship in the Humanities, 10 members of the executive committee for the Australasian Association for Digital Humanities, 9 members of the executive committee for the Association for Computers and the Humanities, 9 members of the executive committee for the European Association for Digital Humanities, and the 4 executive council members in the Canadian Society for Digital Humanities.While I do have great respect for these organizations and many of the people involved with them, the total of absence of any professional statisticians—and in many of the cited examples, lack of scholars with a terminal degree in any technical field—is a problem for a field grounded, at least in part, by the analysis of data.
In the last line of her response “Final Comments,” Da calls for a peer-review process “in which many people,” meaning statisticians and computer scientists, “are brought into peer review.” That is a good place to start but not nearly sufficient. I, and likely many other computationally trained scholars, am already frequently asked to review papers and abstract proposals for the aforementioned journals and professional societies. Da as well has claimed that her Critical Inquiry article was also vetted by a computational reviewer. The actual problem is instead that statisticians need to be involved in computational analyses from the start. To only use computational scholars at the level of peer-review risks falling into the classic trap famously described by Sir Ronald Fisher: consulting a statistician after already having collected data is nothing more than “a post mortem examination.”
To see the potential for working closely with statisticians, one must look no further than Da’s own essay. She critiques the overuse and misinterpretation of term frequencies, latent Dirichlet allocation, and network analysis within computational literary studies. Without a solid background in these methods, however, the article opens itself up to the obvious (at least to a statistician) counterarguments offered in the forum by scholars such as Lauren Klein, Andrew Piper, and Ted Underwood. Had Da cowritten the article with someone with a background in statistics—she even admits that she is “far from being the ideal candidate for assessing this work,” so why she would undertake this task alone in the first place is a mystery—these mistakes could have been avoided and replaced with stronger arguments. As a statistician, I also agree with many of her stated concerns over the particular methods listed in the article.However, the empty critiques of what not to do could and should have been replaced with alternative methods that address some of Da’s concerns over reproducibility and multiple hypothesis testing. These corrections and additions would have been possible if she had heeded her own advice about engaging with statisticians.
My research in computational digital humanities has been a mostly productive and enjoyable experience. I have been fortunate to have colleagues who treat me as an equal within our joint research and I believe this has been the primary reason for the success of these projects. These relationships are unfortunately far from the norm. Collaborations with statisticians and computer scientists are too frequently either unattributed or avoided altogether. The field of DH often sees itself as challenging epistemological constraints towards the study of the humanities and transcending traditional disciplinary boundaries. These lofty goals are attainable only if scholars from other intellectual traditions are fully welcomed into the conversation as equal collaborators.
As a case in point, just last week I had a paper accepted for publication in which we lay out an argument and methodologies for moving beyond word counting methods in DH. See: Arnold, T., Baillier, N., Lissón, P., and Tilton, L. “Beyond lexical frequencies: Using R for text analysis in the digital humanities.” Linguistic Resources and Evaluation. To Appear.
TAYLOR ARNOLD is an assistant professor of statistics at the University of Richmond. He codirects the distant viewing lab with Lauren Tilton, an NEH-funded project that develops computational techniques to analyze visual culture on a large scale. He is the co-author the books Humanities Data in R and Computational Approach to Statistical Learning.
As a computer scientist who has been collaborating in the digital humanities for ten years now, I found Da’s article both well-written and dead on in its arguments about the shallow use of computation. I am teaching a course in text analysis this semester, and I find myself discussing repeatedly with my students the fact that they can computationally find patterns which are almost certainly not causal.
The purpose of computing being insight and not numbers (to quote Richard Hamming), computation in any area that looks like data mining is an iterative process. The first couple of iterations can be used to suggest directions for further study. That further study requires more careful analysis and computation. And at the end one comes back to analysis by scholars to determine if there’s really anything there. This can be especially true of text, more so than with scientific data, because text as data is so inherently messy; many of the most important features of text are almost impossible to quantify statistically and almost impossible to set rules for a priori.
Those first few iterations are the fun 90 percent of the work because new things show up that might only be seen by computation. It’s the next 90 percent of the work that isn’t so much fun and that often doesn’t get done. Da argues that scholars should step back from their perhaps too-easy conclusions and dig deeper. Unlike with much scientific data, we don’t have natural laws and equations to fall back on with which the data must be consistent. Ground truth is much harder to tease out, and skeptical calibration of numerical results is crucial.
Part of Da’s criticism, which seems to have been echoed by one respondent (Piper), is that scholars are perhaps too quick to conclude a “why” for the numbers they observe. Although for the purpose of making things seem more intuitive scientists often speak as if there were a “why,” there is in fact none of that. Physics, as I learned in my freshman class at university, describes “what”; it does not explain “why.” The pull of gravity is 9.8 meters per second per second, as described by Newton’s equations. The empirical scientist will not ask why this is but will use the fact to provide models for physical interactions. It is the job of the theorist to provide a justification for the equations.
There is a need for more of this in the digital humanities. One can perform all kinds of computations (my collaborators and I, for example, have twenty thousand first-year-composition essays collected over several years). But to really provide value to scholarship one needs to frame quantitative questions that might correlate with ideas of scholarly interest, do the computations, calibrate the results, and verify that there is causation behind the results. This can be done and has been done in the digital humanities, but it isn’t as common as it should be, and Da is only pointing out this unfortunate fact.
DUNCAN BUELL is the NCR Professor of Computer Science and Engineering at the University of South Carolina, Columbia.