Translations:BERT Pre-training of Deep Bidirectional Transformers/21/en

The paper also showed that BERT's representations could be used as fixed feature extractors (without fine-tuning) and still achieve strong results, though fine-tuning consistently outperformed the feature-based approach.