COMPARISON BETWEEN HUMANS AND MACHINES ON THE TASK OF ACCENTUATION OF SLOVENE WORDS

The accentuation of Slovene words represents a chal-lenging task for automated solvers since in Slovenian, stress can be located on arbitrary syllables. This pa-per compares the performance of humans, expert-de¯ned rules and computer methods, such as machine learning methods and n-gram Markov models, on this task. We ¯nd that humans tend to accentuate the words correctly, even when they have never heard or seen them before. On the other hand, expert-de¯ned rules for accentua-tion perform quite poorly, achieving worse results than machines. This indicates that humans are good at ac-centuating, but very limited when their knowledge has to be formalized. Therefore, machine methods have to be employed for automatic accentuation of Slovene words.