Bert Models for Russian Sentiment Analysis
Trained Bert-BASE models for russian sentiment analysis. There are PyTorch (transformers) and TensorFlow (deeppavlov) versions available. Models are available for download from Google Drive.
5 dataset was used in this work: Fun, News, RuSentiment, SentiRuEval and TwitterMokoron.
Firstly, 5 separate models were trained on datasets. After that each model was used to predict probability of label to unseen data, probabilities were summarized using model's F1-score as weights and new labels were applied to data. Dataset with new labels from models' predictions are stated as joined dataset later.
Finally, there are 6 labels available for sentiment analysis: positive, negative, neutral, humor, skip and speech.
Models were pretrained from DeepPavlov's Conversational RuBERT on russian sentiment data and after that fine-tuned on joined dataset.
The example of models loading and using is availabe in Use BertClassifier Models for Sentiment Analysis.ipynb.
DistilBert model is also available to download from Google Drive.