Demystifying the Voynich manuscript using computational linguistic techniques

Abstract

The Voynich manuscript is a medieval codex written in undeciphered text by unknown people. In the paper 'How Many Glyphs and How Many Scribes? Digital Paleography and the Voynich Manuscript' by Dr Lisa Fagin Davis (Davis, 2020a), it is proposed that the manuscript is written by five scribes, and a detailed classification is provided. The goal of this research is to provide further evidence to strengthen this proposed classification and to surface any possible misclassification. Towards this end, an experiment was conducted that takes this proposed classification as ground truth and puts it to the test. The data acquired is a transliteration of the manuscript written in the Extensible Voynich Alphabet. It is split into an equal number of pages per scribe; taking into consideration three of the five scribes due to scribes 4 and 5 having much less data. The experiment utilizes stylometric features, namely character bigrams and character trigrams, as features in four different machine learning classifiers; a Deep Neural Network, a Multinomial Naive Bayes classifier, Support Vector Machines and a Decision Tree classifier. This is done in a ten-fold cross-validation where models are trained and predict a scribe for each page. Separately, a k-means unsupervised clustering algorithm is implemented using the same features with k = 3. The results of cross-validation and clustering are compared with the proposal of Dr Fagin Davis.