Improving Results on Russian Sentiment Datasets

TL;DR

This paper evaluates neural network and BERT models on Russian sentiment datasets, finding that BERT-NLI achieves near-human performance, especially with conversational Russian BERT variants.

Contribution

It demonstrates the effectiveness of BERT-NLI for Russian sentiment analysis and compares different BERT variants, highlighting the superiority of conversational BERT.

Findings

BERT-NLI achieves near-human performance on one dataset.

Conversational Russian BERT variants outperform other models.

BERT-based models outperform traditional neural networks.

Abstract

In this study, we test standard neural network architectures (CNN, LSTM, BiLSTM) and recently appeared BERT architectures on previous Russian sentiment evaluation datasets. We compare two variants of Russian BERT and show that for all sentiment tasks in this study the conversational variant of Russian BERT performs better. The best results were achieved by BERT-NLI model, which treats sentiment classification tasks as a natural language inference task. On one of the datasets, this model practically achieves the human level.