TCM-NER

The source code and Dataset of the paper: Qing Ye, Xuan Lai, Chunlei Cheng. Named entity recognition for traditional Chinese medicine with lexical enhancement and span method.

Environment Requirement

torch==1.7.1 
numpy==1.19.5 
transformers==4.26.1 
FastNLP==0.5.0

The Pretrained character embeddings and word embeddings

• Character embeddings: gigaword_chn.all.a2b.uni.ite50.vec
• Bi-gram embeddings: gigaword_chn.all.a2b.bi.ite50.vec

The Pretrained model Chinese-BERT-wwm for embedding layer

Chinese-BERT-wwm

explaination of training parameters

all of the training and hyper parameters are in the file of src/options.py
bert_dir='the path of the Pretrained model Chinese-BERT-wwm'
task_type='crf / span / mrc'  # three methods for decoding
train_epochs                  # the epoch of training, default=10
train_batch_size              # the batch size of training,default=64 
gpu_ids                       # gpu ids to use, -1 for cpu, "0,1,..." for multi gpu

Acknowledgements

• The FLAT model source code.
• The paper of FLAT model: Li X, Yan H, Qiu X, Huang X. J. 2020. FLAT: Chinese NER Using Flat-Lattice Transformer. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 6836-6842
• The detials about FastNLP