Quick follow-up: of the use of distractors in stylometry
This is just a very quick follow-up post to my last post about the Molière-Corneille problem. There, I had noticed that when trying to make the “stylo” script discriminate between Molière and Corneille’s plays, while there was no overlap between the two plays, there also was no clear separating line between the two groups of plays. The supervised method helped with that, and resulted in clear results confirming established knowledge.
This morning, Jan Rybicky mentioned another way of getting a clearer statement on whether or not two groups are separate or not, using the “stylo” script and the consensus tree (which I prefer, because it is more subtle and more visual). This method is that of the “distractor”; it consists of adding a third set of plays to the first two (or more) sets; the trick according to Jan Rybicky is to find a set of relatively similar texts, i.e., essentially similar by genre and time.
Luckily in my case, there is a very good distractor set, namely a number of plays written by Thomas Corneille: there is a set of comedies, and they were published more or less around the same time as the plays by Pierre Corneille and Molière. Adding them to the consensus tree data gives the following result:
[Edit: The addition of a third author pulls apart, in a way, the Molière (red) and Pierre Corneille (green) plays, clearly along the traditionally accepted author lines.] Jan’s comment to this graph was that it worked almost too well! So I also tried it with other playwrights, somewhat more removed from Pierre Corneille and Molière, namely a set of comedies by Scarron and another set of comedies by Regnard; for both of them, the results were indeed less impressive, but still quite useful (one of them is below):
Anyway, the bottom line is that the distractor method is very useful in certain cases, and I’m really glad I was able to add it to my repertory.