The Epistle to Cangrande is one of the most debated documents in the production of the Italian poet Dante Alighieri. For more than a hundred years scholars have been debating over its real paternity, whether it should be considered a work by Dante or a malicious forgery by an unnamed author. In this work, we try to address this philological problem through the methodologies of computational authorship veri cation and machine learning, by training a classi er on a dataset of medieval Latin prose texts and by using a set of authorship-related features. Although the project is still in a preliminary phase, the early results seem to con rm the hypothesis of a forgery.
have been written is not a trivial problem. On the other hand, others (like [2, pp.
280-1] [12, pp. 67-8]) note that there is a lexical coherence spreading through the
entire EpXIII, and an inner cohesive logic. Additionally, [
Given this debated and yet unsolved problem, and in order to gain a fresh perspective over it and thus o er the scholars yet another useful tool for investigation, this project was set to apply the methodologies and technologies of Computational Authorship Veri cation to the mystery of EpXIII. 2
Generally speaking, Authorship Analysis (AA) can be de ned broadly as \any attempt to infer the characteristics of the creator of a piece of linguistic data" [13, p. 238], which includes the author's biographical information (age, gender, mother tongue, etc.) and writing style (use of pronouns, vocabulary richness, etc.), as well as his or her identity. The core of this practice, also known as \stylometry", relies on the idea to identify a certain author not by the artistic value of her/his work, or the meaning of the concepts proposed within it, but via a measure of her/his style. Here, \style" is intended as a summary statistic emerging from one or more numerical features that describe linguistic events present in the written texts, which are believed to remain more or less constant in one's production and conversely vary in noticeable fashion across di erent authors [13, p. 241].
These unique stylistic features are also known as \style markers". This de nition
embraces every kind of textual event, as long as it can be counted (hopefully: easily
counted). It is the researcher's task to identify and extract the most discriminative
features for the type of research s/he is dealing with. In particular, scholars started
experimenting with this practice (well before the age of computers) by employing
a single set of features comparable to the linguistic elements studied in classical
philology, such as the frequencies of certain terms or of word and sentence lengths
[
The values of these stylistic features are collectively used as a simpli ed representations of the respective text, and employed for analyzing its authorship. This may be done via various methods; usually, these methodologies are classi ed as similarity-based or machine learning -based. In particular, in the latter class, a classi er is trained from a number of labelled training examples, using vectors of the chosen features (the style markers) as representations of the texts of interest; this enables the machine to leverage the values of the features in the training examples in order to classify new unlabelled documents into the proper class. These techniques come from the eld of text classi cation and may be tuned to a variety of subtasks: the classes may be literary genres, topics, languages, and so on, depending on the goal of the research. In fact, this is the principle behind many modern applications, from email-spam detecting to authorship identi cation for forensic cases.
In computational AA, the most popular methods still make use of
similaritybased measures or \classical" machine learning algorithms, such as support vector
machines (SVMs) and logistic regression (LR), even if deep learning algorithms
have sometimes proved more accurate. This trend was also con rmed in the PAN
2018 Author Identi cation shared task [
Within this eclectic landscape, the problem of EpXIII is an instance of
Authorship Veri cation (AV), a subtask of AA that aims to design methods and techniques
to determine whether a document of unknown or disputed paternity has been
written by a given candidate author. It is thus di erent from Authorship Attribution,
where the goal is to infer, for a document of unknown or disputed paternity, the
most likely author among a nite set of candidate authors [
AV
We approach the problem of the authorship of EpXIII as a supervised binary classi cation task implemented via linear classi er. In Section 3.1, we give the details of the system employed for the task, while in Section 3.2 we show the training corpus we have assembled; nally, in Section 3.3 we list the results of the related experimentation. 3.1
After a few initial tests with LR and SVMs, we decided to employ the former, since the preliminary experiments on our data had indicated that the two have a similar level of accuracy, and, unlike SVMs, the output of LR admits a probabilistic interpretation, i.e., it can be interpreted as the (\posterior") probability that the document belongs to the class. In particular, after the training phase, the algorithm takes the independent variables (the feature values for the anonymous document) as an input, and computes a value within the logistic function - thus in the interval (0; 1) - as an output. The output is the probability, given the variables, for a certain outcome (belonging to a class, in this case to the Dante class) to happen. See [5, pp. 205-6] for a more complete description of LR.
As already detailed in Section 2, computational methods for AV represent the
textual document as a vector of variables representing some linguistic
phenomena. To this aim, we have selected a combination of various feature types, since
this approach usually assures better performances while dealing with multivariate
methods [
As already explained in Section 1, EpXIII consists of two sections that are very distinct for purpose and style; moreover, many scholars theorize a di erent hand for each of them. Hence, it seemed imperative to follow the traditional division and split the AV problem into two sub-problems, training a speci c classi er for the paragraphs 1-13 (from now on: EpXIII(I)) and another one for the remaining paragraphs 14-90 (from now on: EpXIII(II)).
In order to o er to the classi er an adequate representation of the two authorial classes (Dante and NotDante), we created two corpora of medieval Latin texts by collecting documents which can be considered, linguistically and stylistically speaking, close to the related portion of EpXIII. Understandably, the positive class (Dante) is represented by the works unquestionably ascribed to Dante: the other 12 letters for EpXIII(I), De Vulgari Eloquentia and Monarchia for EpXIII(II). Conversely, for the negative class (NotDante), two sets of texts of coeval Latin authors have been assembled: a number of epistles from various authors (EpXIII(I)) and a collection of di erent textual works, mostly literary comments and treatises (EpXIII(II)). All the documents are dated between the 13th and the 15th century. Additionally, we subject the training documents to a segmentation policy in order to further increase the number of training samples. Using a large number of training segments for each author when only few long texts are available is a common practice in the eld [17, p. 514]. The nal result of this procedure is shown in Table 1. In order to determine the degree of reliability of the classi er, and hence establish the trustworthiness of the classi cation hypotheses it generates, we subject the algorithm to a \leave-one-out" validation test; more speci cally, we create an AV classi er for each author that has more than one positive sample in the corpus, setting the author as the positive class. In particular, as to recreate the condition of the actual classi cation of EpXIII(I) and EpXIII(II) and to avoid overlapping between test and training samples, we only test the full original documents and, when the document t is used as test, all the segments derived from t are excluded from the training. The results of the validation test are shown in Table 2; we use Macro-F1 and Micro-F1 as evaluation measures. As it can be seen, the classi ers that have been trained, and especially the one for EpXIII(I), obtain a good level of accuracy, in line with other state-of-the-art methods. Finally, the classi cation hypotheses are shown in Table 3, along with the F1 values for the class Dante. As it can be seen, the classi ers consider both portions of EpXIII written by someone other than Dante, with a slightly lower probability, and thus higher con dence in the negative attribution, for EpXIII(I). Considering the speci c accuracy level obtained in the validation, while the value for EpXIII(I) is satisfying, the same cannot be said about the value for EpXIII(II); the classi er is penalized for not attributing Monarchia to Dante, while it correctly classi es 26 out of the 28 negative samples.
EpXIII(I) EpXIII(II)
NotDante NotDante Henceforth, the hypotheses computed seem to align with the theory that the entire EpXIII was the production of a malicious forger. Nevertheless, the evaluation presented here should not be considered conclusive: in future developments we plan to de ne some improvements to the system, such as an expansion of the training corpus and of the feature set exploited, as well as experimenting with some di erent visualization techniques. 4
In this project, we explore a case of AV applied to Latin prose, in order to infer the real author of the Epistle to Cangrande, a letter of dubious Dantean paternity, via computational and quantitative methods.
The classi cation hypotheses we reach seem to con rm the forgery thesis. However, the current stage of the research is still preliminary. It is important to underline that our goal is not to reach some de nitive answer for the philological dilemma; especially for this kind of problems, where the truth is debated even among specialists, the hypothesis given by a classi cation algorithm can not, and should not aim to, be nal. Nevertheless, we are eager to explore other methodological possibilities that could determine an improvement in the system accuracy, and generally create a more accessible and solid tool that can aid the study of the researcher, not replace it.
Acknowledgments. The research described in this paper is part of the work carried
out by the author for her MSc thesis in Digital Humanities at the University of
Pisa (supervisors: Alejandro Moreo, Fabrizio Sebastiani, Mirko Tavoni), discussed
in February 2019 and also reported in [