In this work we try to classify 5,063 blog texts written by 68 authors using XLM Roberta
Large Xnli model (from Facebook AI) and a vocabulary of Russian antonyms. We find that the zero-shot
classification does not provide sufficiently good results for this task. We suggest that maybe the
benchmarks need additions if a model can provide good results on the main benchmarks and not good
results in other tasks.