This paper describes a statistical methodology for a diachronic study on a large corpus (a collection of publications, written from the second decade of the 19th century in two countries, Romania and Republic of Moldova, known as Bessarabia). The aim of this work is to analyse the lexical evolution of words in all four regions using a machine learning approach to identify the patterns that govern language changes. Basically, it was developed a mechanism for automatic correlation of different forms of the same words in order to train a statistical model on a list of known word-to-word correlations between lexicons.