Higher Criticism for Discriminating Word-Frequency Tables and Testing Authorship

We have a new document of unknown authorship; we would like to determine its author. We have several corpora of documents, each of homogeneous authorship, and we believe the unknown author of the new document is represented among our corpora. The unprecedented abundance and availability of text data in our age generates manyauthorship attribution problems of this form. Existing approaches for such problems usually construct a set of handcrafted features to discriminate between potential candidate authors [1, 2, 3, 4]. Typically, these features originate from linguistic heuristics, such as rate of usage of certain words and length of sentences, and are often first constructed by trial and error, or based on domain expertise or historical tradition.