Improve model performance

[dveselov]

We can create a large automatically annotated corpus as was done with natasha/nerus and train fasttext classifier on in.
It looks like deeppavlov BERT model has better accuracy than we have and it was trained on rusentiment too, so we can use it for knowledge distillation.