russian classification sentiment multiclass

Sentiment model based on rubert-base-cased-conversational

This model was initialized with rubert-base-cased-conversational weights and trained on a batch of datasets collected by Smetanin, using the same training sampling presented in this wonderful work. This approach allows for a uniform distribution among different datasets and three classes of sentiment labels: negative, neutral, and positive. Datasets were prepared by David Dale and are hosted here.

I chose rubert-base-cased-conversational weights because, according to Smetanin's work, this model ranks first among all other multilingual and popular Russian language models with BERT base architecture.

Training and Testing Details

This model was trained and tested using the code and hyperparameters from the rubert-tiny-sentiment-balanced work.

Labels

There are only three labels: negative - 0, neutral - 1, positive - 2

Results

It outperforms rubert-tiny-sentiment-balanced on four datasets, underperforms on one (linis), and has the same performance on mokoron and rureviews. See this for the comparison.

Source Macro F1
SentiRuEval2016_banks 0.88
SentiRuEval2016_tele 0.79
kaggle_news 0.73
linis 0.46
mokoron 0.98
rureviews 0.77
rusentiment 0.74