There are two distinct things going on for me in this lab. The first is a simple methodological survey of topic modeling, a how-to. This is definitely the smaller part of the labor but is important to talk about nonetheless. The process was naturally prescriptive and I plan to take more time in the future to work with different tools to get a more robust sense of the ways in which topic modeling can be done. Mallet was a good start but I want to see what other interfaces I can make use of. Returning to the lab itself, topic modeling is a simple process to execute but a tough one to master. My model is simple and I did not play with the commands much at all to tweak it. I went about things this way because I was afraid of manipulating the wrong variable and getting a useless model. This fear is indicative of an inherent difficulty with topic modeling: I need it to make sense. What comes out of Mallet needed to follow some unspoken syntax so that meaning could be derived from it. This fed an ill-defined anxiety as I executed my model that everything would fall apart if I did not stay within the parameters of the established code. In other words, what I did was not all that interesting but why I did it is. I find myself wondering what a “proper” model should be and whether or not I should care about that. Legibility actually does not enter into this lab at all and neither do negative consequences for experimentation. Still, I was hesitant to produce schlock. I suppose this is a roundabout way to make a case for failure, especially where topic modelling is concerned. The less one is concerned with accuracy or validity (and this may be applicable to all theoretical matters in literary studies) the more one can focus on sharpening a method rather than a claim.
As for my personal experience with method in this lab(the second part), my model told me some interesting things about domesticity that may yield something if pushed a bit more. I was able to see a strong relationship between the words “home,” “money,” and “London.” Too, the word “woman” appeared with frequently with words like “house” and “young.” I should say at this point that my corpus is composed of late 19th and early 20th century detective fiction. This helps to make a bit of sense out of the hypothesis that these connections inform a reading of the corpus as reinforcing social norms while offering a tantalizing window into the scandal of crime and otherness.
There were a few hiccups that helped me to see a few problems with that initial corpus. I mixed novels with short stories and got a great deal more influence out of those words that were prevalent in the novels simply because they appeared more often by default. “Necklace” and “Baskervilles” were over-represented because I added the longer texts Hound of the Baskervilles and The Moonstone to the corpus without realizing how much more voluminous they were in comparison to, say, “The Purloined Letter.” So even though I was reluctant to make mistakes with the model, fortunately I built that initial corpus without anything in mind and learned something about corpus-building that I would not have otherwise.
I find myself in a strange position this week where I regret not failing more. This course (and this is a bit tangential) is structured in such a way that my mistakes are more valuable than my successes. I do not know if this bears on pedagogy writ large or merely speaks to the position of digital methodologies in the contemporary theoretical scene. I am also not so naïve as to think constant failure is the object of my work or anyone else’s. Still, I wonder if these tools lead us to think about failure differently.