Automatic Morphological Analysis for Russian: a Comparative Study

In this paper we present a comparison of ten systems for automatic morphological analysis: TreeTagger, TnT, HunPos, Lapos, Citar, Morfette, Mystem, Pymorhy, Stanford POS tagger and SVMTool. Different training and tagging approaches are discussed together with the strengths and weaknesses of each system. Probabilistic taggers were trained and tested on the Russian National Disambiguated Corpus and achieved accuracy scores as high as 96,94% on POS tags and 92,56% on the whole tagset. However, most of the existing taggers cannot resolve various cases of morphological ambiguity and show a better performance for morphologically rich languages. We believe that the detailed examination of errors caused by homonymy can help to solve the disambiguation problem and to improve tagging results.