Computational authorship studies are an increasingly popular topic for research among specialists sopra both calcolatore elettronico science and the humanities

It can be considered a form of style-based document authentication (Echtheitskritik), which has valuable applications that extend well beyond the domain of literary analysis, preciso, for instance, the domain of forensic sciences. According sicuro Stamatatos’s 2009 survey of the field, ‘[t]he main ispirazione behind statistically or computationally-supported authorship attribution is that by measuring some textual features we can distinguish between texts written by different authors.’22 22 Anche. Stamatatos, ‘Verso survey’ (n. 14, above) 538. This basic assumption implies that it should be possible to assess, for any new unseen document, whether or not it was written by other authors for whom we have texts available. Nowadays computational authorship studies are often considered a subfield of stylometry sopra the digital humanities, the broader computational study of the writing style of texts.23 23 D. Holmes, ‘The evolution of stylometry durante humanities scholarship’, LLC 13 (1998) 111–17.

While stylometry has verso rich history, dating back onesto at least the nineteenth century, it is clear that it received its most important impetus only con the past two or three decades, stimulated by the rise of (personal) computing and the increased availability of large bodies of text mediante electronic form. Apart from the influential, yet more conventional, statistical analyses carried out by pioneers such as Mosteller and Wallace or John Burrows well before the 1990s, an influential approach sopra authorship studies has been preciso approach the attribution of anonymous texts as verso ‘text categorization’ problem.24 24 Mosteller and Wallace, Inference and disputed authorship (n. 4, above) and J. Burrows, Computation into criticism: a study of Jane Austen’s novels (Oxford 1987). Heavily influenced by parallel research per computer science, the timore was preciso optimize per statistical classifier on example texts by per number of available candidate authors, much like a spam filter nowadays is still trained on manually annotated emails sicuro learn how onesto distinguish between ‘junk’ email and normal messages.25 25 F. Sebastiani, ‘Machine learning sopra automated text categorisation’, ACM Computer Surveys 34 (2002) 1–47. After preparazione such per classifier on this example momento, the classifier could then be used onesto categorize or classify anonymous text as belonging esatto one of the training authors’ oeuvres.

It resembles verso police lineup, mediante which the correct author of an anonymous text has puro be singled out from per series of available candidate authors for whom reference or ‘training’ material is available

This text categorization setup is commonly known as ‘authorship attribution’.26 26 The following paragraph heavily draws on M. Koppel and Y. Winter, ‘Determining if two documents are written by the same author’, JASIST 65 (2014) 178–187. For a number of years, practitioners of stylometry have che razza di preciso acknowledge the limitations of authorship attribution, because it necessarily assumes that the correct target author is indeed included sopra the servizio of candidates. In many real-world cases, this problematic assumption cannot possibly be made, because the serie of relevant candidates is difficult or impossible preciso establish beforehand. Because of this, the setup of authorship verification has recently been introduced as per new framework: here, the task is to verify whether or not an anonymous document was written by one or several of per series of candidate authors. Durante some sense, authorship verification redefines the text categorization problem by adding an additional category label: ‘None of the above.’

In the present context, it should be emphasized that the problem posed by the HA is verso ‘vanilla’ example of per problem per authorship verification: while the campione indeed contains a number of (auto-) attributions, the veracity of all of these has been questioned sopra previous scholarship

Verification is hence an increasingly common experimental setup sopra authorship studies, and is the topic of a dedicated track per the yearly PAN competition, an annual competition on finding computational solutions puro issues durante present-day textual forensics, mostly related puro the detection of plagiarism, authorship, and affable software misuse (such as grooming or Wikipedia vandalism).27 27 The competition’s website is pan.webis.de. The most recent survey of an authorship verification track is: Anche. Stamatatos et al., ‘Overview of the author identification task at PAN 2015′ per Working Taccuino Papers of the CLEF 2015 Evaluation Labs, di nuovo. L. Cappellato et al. (2015). Generally speaking, authorship verification is per more generic problem than authorship attribution – i.addirittura. every attribution problem could , mediante principle, be cast as verso verification problem – but it has also proven esatto be more challenging. In our experiments, we have therefore attempted to radically minimize any assumptions on our part as puro the authorial provenance of the texts sopra the HA. For each piece of text analysed below, we propose sicuro independently assess the probability that it was written by one of the (alleged) individual authors identified sopra the corpus.