In this paper we focus on syntactic annotation consistency within Universal Dependencies (UD) treebanks for Russian: UD_Russian-SynTagRus, UD_Russian-GSD, UD_Russian-Taiga, and UD_Russian-PUD. We describe the four treebanks, their distinctive features and development.
In order to test and improve consistency within the treebanks, we reconsidered the experiments by Martínez Alonso and Zeman; our parsing experiments were conducted using a state-of-the-art parser that took part in the CoNLL 2017 Shared Task. We analyze error classes in functional and content relations and discuss a method to separate the errors induced by annotation inconsistency and those caused by syntactic complexity and other factors.