Modeli za automatsku morfološku fleksiju srpskog i hrvatskog jezika na osnovu morfoloških leksikona srLex i hrLex

Open-source machine-readable morphological lexicons are useful for morphosyntactic tagging of corpora and represent a crucial step toward compiling modern digital dictionary databases. In the paper, we present the first step toward extending the functionalities of Pregibalnik, a custom developed open-source tool for Slovene lexicon expansion, to cover Serbian and Croatian and help automatically expand the lexicons with new entries. We describe the process of extraction of morphological patterns from the hrLex and srLex inflectional lexicons of Croatian and Serbian, as well as a robust process of feature selection based on ending word parts. The features are used to develop a series of machine-learning models to predict morphological patterns for Croatian and Serbian lexemes, achieving an average F1-micro score of 0.85 (depending on lexeme type). This also helps identify potential inconsistencies within the current versions of the lexicons. The extracted patterns and models are available under a Creative Commons CC-BY 4.0 license.