Using machine learning for language and structure annotation in an 18th century dictionary
Authors: Petra Bago, Nikola Ljubešić
The accessibility of digitized historical texts is increasing, which, consequently, has resulted in a growing interest in applying machine learning methods to enrich this type of content. The need for applying machine learning is even greater than in modern texts given the high level of inconsistency in historical texts even within the same document. In this paper we investigate the application of a supervised structural machine learning method on language and structure annotation of 18th century dictionary entries. Our research is conducted on the first volume of a trilingual dictionary ‘Dizionario italiano–latino–illirico’ (Italian–Latin–Croatian Dictionary) compiled by Ardellio della Bella and printed in Dubrovnik in 1785. We assume that by using this method, we can significantly reduce time for manual annotation and simplify the process for the annotators. We reach accuracy of approximately 98% for language annotation and around 96% for structure annotation. A final experiment on the time gain obtained by pre-annotating the data shows that only correcting the generated labels is roughly five times faster than full manual annotation.
Reference: In Kosem, I., Jakubiček, M., Kallas, J., Krek, S. (eds.) Electronic lexicography in the 21st century: linking lexical data in the digital age. Proceedings of the eLex 2015 conference, 11-13 August 2015, Herstmonceux Castle, United Kingdom. Ljubljana/Brighton: Trojina, Institute for Applied Slovene Studies/Lexical Computing Ltd., pp. 427-442.