Identifying Latin Authors through Maximum-Likelihood Dirichlet Inference: A Contribution to Model-Based Stylometry

Shagi/Steps. 2021. Vol. 7. No. 1. P. 183–198.

Nikolaev D., Shumilin M.

The last two decades saw a dramatic increase in the number of papers published on the subject of stylometry, which is often narrowly understood as the task of identification of the author of a particular text fragment based on its stylistic properties. We present a new lightweight algorithm for stylometric identification of authors of Latin prose texts based on Burrows’s Delta, computed over relative frequencies of 244 manually selected genre and topic neutral words, and the Dirichlet distribution, whose parameters we estimate using an iterative maximum-likelihood algorithm. In order to demonstrate the effectiveness of the method, we present a case study of 3000-word fragments of texts by 36 classical and medieval authors and show that our method performs on par with Random Forest, a powerful general-purpose classification algorithm. We provide summary statistics of our algorithm’s performance together with confusion matrices demonstrating pairwise discriminability of texts by different authors. The advantages of our method are that it is very simple to implement, very quick to train and do inference with, and that it is very interpretable since it is a model-based algorithm: precision of the fitted Dirichlet distributions directly corresponds to the stylistic homogeneity of the texts by different authors. This makes it possible to use the algorithm as a general research tool in Latin stylistics.

Keywords: machine learning, Latin literature, stylometry, text attribution, random forest, Dirichlet distribution, Burrow's Delta, stylistic analysis