This paper presents a system for automatic labelling of natural language texts according to a more or less detailed system of linguistic categories (grammatical, syntactical, etc.). A Markovian model is used to predict the label of each word of the unknown text. Several assumptions and restrictions improve the computational efficiency with a small decrease of the performance of the system. This has been measured by labelling 120.000 words of Greek newspaper texts with grammatical labels and proved to be satisfactory.