Let x represent an unknown document and let y represent a random target author’s stylistic ‘profile’. During one hundred iterations, it will randomly select (a) fifty verso cent of the available stylistic features available (addirittura.g. word frequencies) and (b) thirty distractor authors, or ‘impostors’ from per pool of similar texts. Con each iteration, the GI will compute whether quantita is closer sicuro y than preciso any of the profiles by the thirty impostors, given the random selection of stylistic features in that iteration. Instead of basing the verification of the direct (first-order) distance between quantitativo and y, the GI proposes to primato the proportion of iterations mediante which x was indeed closer esatto y than esatto one of the distractors sampled. This proportion can be considered per second-order metric and will automatically be verso probability between zero and one, indicating the robustness of the identification of the authors of quantitativo and y. Our previous work has already demonstrated that the GI system produces excellent verification results for classical Latin prose.31 31 Padrino the setup visitatori muzmatch sopra Stover, et al, ‘Computational authorship verification method’ (n. 27, above). Our verification code is publicly available from the following repository: This code is described in: M. Kestemont et al. ‘Authenticating the writings’ (n. 29, above).
For modern documents, Koppel and Winter were even able sicuro report encouraging scores for document sizes as small as 500 words
We have applied a generic implementation of the GI preciso the HA as follows: we split the individual lives into consecutive samples of 1000 words (i.ancora. space-free strings of alphabetic characters), after removing all punctuation.32 32 Previous research (see the publications mentioned con the previous two taccuino) suggests that 1,000 words is a reasonable document size in this context. Each of these samples was analysed individually by pairing it with the profile of one of the HA’s six alleged authors, including the profile consisting of the rest of the samples from its own text. We represented the sample (the ‘anonymous’ document) by verso vector comprising the correspondante frequencies of the 10,000 most frequent tokens sopra the entire HA. For each author’s profile, we did the same, although the profile’s vector comprises the average divisee frequency of the 10,000 words. Thus, the profiles would be the so-called ‘mean centroid’ of all individual document vectors for verso particular author (excluding, of course, the current anonymous document).33 33 Koppel and Seidman, ‘Automatically identifying’ (n. 30, above). Note that the use of a single centroid per author aims puro ritornato, at least partially, the skewed nature of our tempo, since some authors are much more strongly represented in the corpus or sostrato pool than others. If we were not using centroids but mere text segments, they would have been automaticallysampled more frequently than others during the imposter bootstrapping.
Preciso the left, a clustering has been added on apice of the rows, reflecting which groups of samples behave similarly
Next, we ran the verification approach. During one hundred iterations, we would randomly select 5,000 of the available word frequencies. We would also randomly sample thirty impostors from a large ‘impostor pool’ of documents by Latin authors, including historical writers such as Suetonius and Livy.34 34 See Appendix 2 for the authors sampled. The pool of impostor texts can be inspected in the code repository for this paper. Sopra each iteration, we would check whether the anonymous document was closer preciso the current author’s profile than to any of the impostors sampled. In this study, we use the ‘minmax’ metric, which was recently introduced sopra the context of the GI framework.35 35 See Koppel and Winter, ‘Determining if two documents’ (n. 26, above). For each combination of an anonymous text and one of the six target authors’ profiles, we would superiorita the proportion of iterations (i.addirittura. per probability between niente and one) durante which the anonymous document would indeed be attributed preciso the target author. The resulting probability table is given durante full per the appendix onesto this paper. Although we present verso more detailed conciliabule of this momento below, we have added Figure 1 below as an intuitive visualization of the overall results of this approach. This is verso heatmap visualisation of the result of the GI algorithm for 1,000 word samples from the lives con the HA. Cell values (darker colours mean higher values) represent the probability of each sample being attributed preciso one of the alleged HA authors, rather than an imposter from per random selection of distractors.