Score-based Likelihood Ratios Using Stylometric Text Embeddings

We consider the problem setting in which we have two sets of texts in digital form and would like to quantify our beliefs that the two sets of texts were written by the same author versus by two different authors. Motivated by problems in digital forensics, the sets of texts could be composed primarily of short-form messages, and texts by the same author may be about vastly different topics. To this end, we focus on user-specific stylometric aspects of the texts that are consistent across an author’s writings and are invariant to topics. Recent work in machine learning has sought to learn a mapping from input texts to output a vector representation intended to capture such stylometric features. In this work, we investigate the use of such stylometric text embeddings to construct a score-based likelihood ratio (SLR), an increasingly popular way of quantifying evidence in forensics. We present the results of SLR experiments using recently proposed stylometric embeddings from machine learning applied to real-world datasets relevant to digital forensics.