RuHateBe
A Russian hate-speech benchmark for dialogue systems, built to measure how readily conversational models produce abusive replies, broken down by the group being targeted rather than averaged into one safety score.
Hanna Shcharbakova · additional projects
Additional Projects
2022
2020
vkToxic
A Python library for sentiment analysis of comments from VKontakte communities, pairing automated collection with classification tuned for communities at both ends of the toxicity range.
Russian SuperGLUE
Metaphor-based diagnostic tasks for the Russian SuperGLUE benchmark — test methodology and annotation for probing whether models understand figurative language.
2017–2018
Universal Dependencies
Syntactic annotation and morphological tagging for the Belarusian treebank: automated parsing corrected by hand, against the Universal Dependencies guidelines.