Fine-Tuning LSTM-CRF model with HuggingFace's Transformers

On top of the previous version, we add another option to fine-tune BERT/Roberta/etc-based LSTM-CRF model using HuggingFace's Transformers package.

Our latest benchmark results show that we obtained SOTA results on both CoNLL-2003 and OntoNotes datasets. And we also show that Roberta-CRF is better than BERT-CRF.