Short samples in authorship attribution: a new approach

The question of minimal sample size is one of the most important issues in stylometry and nontraditional authorship attribution. In the last decade or so, a few studies concerning different aspects of scalability in stylometry have been published (Zhao and Zobel, 2005; Hirst and Feiguina, 2007; Stamatatos, 2008; Koppel et al., 2009; Mikros, 2009; Luyckx and Daelemans, 2011), but the question has not been answered comprehensively. In his recent study, Eder proposed a systematic approach to solve the problem in a series of experiments, claiming that a sample should have at least 5,000 running words to be attributable (Eder, 2015). The above studies (and many other as well) tacitly assume that there exists a certain amount of linguistic data that allows for reliable authorial recognition, and the real problem at stake is to determine that very value. However, one can assume that the authorial fingerprint is not distributed evenly in a collection of texts. Just the contrary, many experiments seem to suggest that the authorial voice is sometimes overshadowed by other signals, such as genre, gender, chronology, or translation. Some authors, say Chandler, should be easily attributable, while some other authors, say Virginia Woolf, will probably have their fingerprint somewhat hidden. Moreover, authorship attribution is ultimately a matter of context: telling apart Hemingway and Dickens will always be easier than distinguishing the Bronte sisters. On theoretical grounds, then, the minimal sample size can not be determined once and forever for the entire corpus, but may be different for different texts in the corpus.