Introduction
BERT has achieved great success in the field of natural language processing (NLP). Using unlabeled data sets for training, large-scale models that can learn complex language representations can be obtained. Then, we can apply similar research methods to chemical representations, especially SMILES sequences:
Self-supervised learning task
1. Masked language modeling (MASKEDLM)
The normative task proposed by BERT is to predict the true identity of the mask by training the model. Use the cross-entropy loss between the sequence output and the input mask to optimize the task.
2.SMILE