Word embeddings for detection of verbal idioms in Slovene

Word embeddings map words to a high dimensional vector space, where words with similar meanings have similar vectors. We analyzed the problem of automatic identification of verbal idioms in Slovene using features built from embeddings of single words and groups of words. For this purpose, we built two data sets that contain verbal idioms and random word groups described with corresponding features. Using these data sets we evaluated the classification of verbal idioms with support vector machines, random forests, and logistic regression. All three methods were successful, the best being random forests. Due to large computational time and limitation to only identify groups of words with precomputed word embeddings the approach requires further improvements to be practically useful