The Computational Literary Studies Debate: Thoughts from the Fall 2019 Class, Writing in Digital Humanities

Is There Clear Evidence That CLS is Detrimental?

Digital Humanities has grown tremendously in the past decade. At its core, digital humanities is the crossroad of computation and other disciplines such as history and literature. The hope of this is to enhance the understanding of these texts and lead to new discoveries based on physical evidence.

Computational Literacy Studies (CLS) strives to provide coherent analysis of literature text through computation. This is in the form of word counting which would then correspond to not only the basic analysis of the text, but the emotion of the text as well. The use of word counting raises the question can CLS truly capture literature objectively and accurately through numbers. While CLS goes beyond word counting, its basic function is to count words which is an ineffective way to analyze advanced literature. 

“No matter how fancy the statistical transformations, CLS papers make arguments based on the number of times x words or gram appears. CLS processing and visualization of data are not interpretations and readings in their own right.To believe that is to mistake basic data work that may or may not lead up to a good interpretation…” (Nan Da 6).

Nan Da, an author, explored how accurate CLS is in providing a conclusion that is in depth and valid towards a literary text. In “The Computational Case against Computational Literary Studies,” by Nan Da there are criticisms against CLS and how it misrepresents literature, because it lacks the ability to provide accurate evidence towards a claim. For example, word vectorization is used to link a hypothesis to a causation, which Nan Da believed was fundamentally wrong. This concept of word vectorization was explored in Mark Algee-Hewitt’s experiment on bigrams. A bigram is the likelihood one word will be followed by another word. How often bigrams appear indicate some type of connectivity which leads to a hypothesis. For example, the word “I” followed by the word “love” indicates some type of affection. However, a hypothesis needs statistical tools to show strong correlation, but fundamental literacy principles such as word counting can’t viably link this relationship. Since this relationship can’t be linked viably, the causation of the hypothesis isn’t defined. If there is an undefined hypothesis, then the experiment or analysis is already flawed and therefore will yield either wrong or deceiving conclusions. 

Furthermore, Da argues that data mining text labs are given disproportionate amount of funds in respect to what they offer or produce (3). In reality, a smartphone can handle the computing power necessary to do most of the computations, which raised the question: why is there a need for entire text mining labs? Text mining labs are naturally extremely expensive. Nan Da ultimately believes funding for CLS needs to be cut completely, because it is wasting money when literature exploration funding is already extremely limited. Since CLS at its core is word counting the need for text mining labs is even more unnecessary. Her personal biases against CLS are highlighted and because of this, the lines between a research paper and an editorial are blurred. On the flip side, Da also states that her essay does not claim that numbers are unethical or even the science of text mining on the surface is inherently wrong, but her problems arise from deceiving arguments CLS often makes.

Consequently, Da says, CLS often provides deceiving information, because it only utilizes word frequencies, disregarding syntax, position, and background (11). CLS, Nan Da argues, attempts to make claims without context which leads to misinterpreted conclusions. For example, Paul Vierthaler’s work on different types of Chinese writings claimed that the two genres of Chinese writings, historical apocrypha and fiction, are not as similar as once thought by historians. He used word frequencies to highlight this by taking books and splitting them into 10 thousand character chunks and then took the one thousand most used Chinese characters in that chunk (Da 15). Vierthaler then makes a graph to not only emphasize his point, but to make the claim that historical apocrypha is closer to official history due to the similarities between the languages.

To strengthen his view, Vierthaler points to commonalities between theme and plot. However, the differentiation comes from the content not by the formal language. Two pieces of writing can have very similar language use, but can lead to two different conclusions. Therefore, by using word frequencies, the conclusion will be misconstrued. 

In “Statistical Modeling: The Two Cultures,” Leo Breiman argues statistics have rapidly developed false narratives. Goodness of fit tests, which are the discrepancies between the outcome given and the outcome expected, are utilized to make false narratives (Breiman 4). For example, there was a study at a university to measure if there is discrimination in the salary of the faculty working at the university. Salary was used as the response variable while 25 other variables which characterized academic performance, evaluation, journals published, etc were also used (Breiman 4). Breiman believed there are so many variables that one can’t deduce the conclusion to a binary yes or no. This lead to him deriving that these analysis are deficient, because they focus on the model and not the problem itself, especially when the model itself has interference with many variables.

Similarly, Breiman thought that the data gathered from CLS for the most part was deficient, because the lack of repetition. He used the tree rate example to emphasize this point. Breiman worked on a project that analyzed delay in criminal court cases at the state level. Specifically in Colorado, which had excellent computerized court data systems. The dependent variable used was the time from arraignment to sentencing. Everything else was a predictor variable. A large “decision tree” was formed which just split up Colorado into districts. Breiman then presented his findings to Colorado judges. It was shown that district N had a long delay time relative to the other districts (Breiman 9). The problem with this study, in Breiman’s eyes, is the fact that it was not replicated to ensure correct results.
(Picture is suppose to be here)

Instead of using a “tree” as the foundation for the results, Breiman believed having a “forest” or multiple tests on the same thing, is the best course of action to ensure the most reliable results. This leads to the conclusion for CLS to add reliable results, it will need to process the data numerous times over. However, Breiman believes this is currently not happening, which is extremely problematic. CLS has to be held to a high reliability standard as it utilizes statistics as a tool for analysis. This also proves the point that more money and time will have to be spent on computations.

Both Nan Da and Breiman have the same end conclusion on CLS and how effective its applications are. Breiman comes from a very statistical heavy background, so his opinions are extremely statistics oriented. Nan Da goes into less technical criticism, but also makes claims across a larger board. In a figurative sense the problem with CLS, according to both Nan Da and Breiman, is that CLS only tool is a hammer which makes every problem look like nails. However, this is problematic, because every problem is not a nail and therefore another tool needs to be added to CLS to be effective. Da and Breiman would change their minds about CLS if more tools were added to the toolbox. The make up of CLS currently is not only ineffective, but a waste of time and money. Money and time should go into CLS research to help make it better by addressing the issues as stated instead of continuing CLS applications as they are right now. 

For CLS to fully become transformative on literature, there are major gaps that need to be addressed. Statistics need to match what it is trying to measure instead of producing deceiving conclusions. Word frequencies can reveal information about a literary text, but context is also tremendously important in analyzing the text fully. If these gaps are addressed then the debate of CLS will change entirely. Right now, CLS is problematic and unnecessary, especially once the difficulty of finding funding for literature research is considered. Without improving CLS’s functionalities, analysis of literature will lack the in depth scrutiny that it requires. Furthermore, CLS just repeats what's already been established in literature text or just doesn’t enhance the understanding of that literature text making it unnecessary.

Work Cited

Da, Nan Z. “The Computational Case against Computational Literary Studies.” Critical Inquiry

vol. 45, no. 3, 2019, pp. 601–639., doi:10.1086/702594.

Breiman, Leo. “Statistical Modeling: The Two Cultures.” Vol.16, No. 3, Statistical Science Institute of Mathematical Institute, 2001,


This page has paths:

This page references: