Conceptual

BERT Bidirectional Transformer Encoder Pretraining for Language Representation

A pre-training method for language representation in which a multi-layer bidirectional Transformer encoder is trained on large unlabeled text using masked-language-modeling (predicting randomly masked tokens from both left and right context) and next-sentence prediction. The resulting contextual embeddings are then fine-tuned with a task-specific head for downstream NLP tasks such as classification.