Authorship verification for Cultural Heritage: the case of the apocryphal Quixote

"El segundo tomo del ingenioso hidalgo Don Quixote de la Mancha" is an apocryphal work published in 1614 under the pseudonym of Alonso Ferna ́ndez de Avellaneda, and intended as a continuation of the first part of Miguel de Cer- vantes’ immortal novel. Over the centuries, various hypotheses have been put forward as to its authorship. However, there is still no agreement among scholars on who the author of this “aprocryphal Quixote” might be. In this dissertation, we propose the application of computational authorship identification methods based on machine learning to the problem of identifying the author of the aprocryphal Quixote. These methods are quantitative, and based on the automatic extraction and analysis of the stylistic (in the sense of stylometry) elements of a text. In particular, our research focuses on the develop- ment of authorship verification models, i.e., models (implemented as automatic binary classifiers) whose task is to confirm or reject the hypothesis that a text has been written by a given candidate, and authorship attribution models, i.e., models (implemented as automatic single-label multiclass classifiers) whose task is to determine the most likely author of a text from a closed set of candidates. Contrary to the hypothesis recently formulated by several scholars, the mod- els we have developed seem to exclude that Jero ́nimo de Pasamonte (a former comrade-in-arms of Cervantes) might be the author of the apocryphal Quixote. Indeed, our models seem to exclude that any of the 13 Spanish authors we have considered, all of them indicated as possible candidates by Cervantine scholars, is Alonso Ferna ́ndez de Avellaneda. An interesting aspect that emerges from our analysis is the outcome of the authorship attribution model, which indicates Miguel de Cervantes himself as the most likely author among the 13 candidates considered. This result suggests that the apocryphal Quixote imitates Cervantes’ style very effectively, since it misleads an automatic attribution model that, as we show in the dissertation, proved 100% accurate in previous demanding experiments. However, as argued above, Avellaneda’s style is not similar enough to Cervantes to deceive our equally accurate (as we also show) verification model.