POS Tagging for Chinese

Model : CharLSTM + LSTM + CRF ​ Elmo + LSTM + CRF ​ Parser + CharLSTM + LSTM + CRF

Pretrained Word Embedding : glove.6B.100d.txt

NER Data : CoNLL03 Chunking Data : CoNLL00 POS Data : WSJ

Other resources: Elmo : produce elmo representation on all datasets, and save on the disk Biaffine Parser : produce biaffine parser's LSTMs output on all datasets, and save on the disk

requirements

python >= 3.6.3
pytorch = 0.4.1

running

mkdir save # or define other path to save models and vocabs
python train.py --pre_emb --task=ner --gpu=0 # choose task and if use gpu and pretrain embedding
python train_parser --pre_emb --task=ner --gpu=0
python train_elmo.py --pre_emb --task=

results

Task:NER

Data:CoNLL03

Pretrained Embedding:glove.6B.100d.txt.

model	dev	test	Iter
CharLSTM+LSTM+CRF	94.47%	91.10%	18/29
Elmo+LSTM+CRF	95.56%	92.12%	25/36
Parser+CharLSTM+LSTM+CRF	94.87%	90.87%	35/46

Task:Chunking

Data:CoNLL00

Pretrained Embedding:glove.6B.100d.txt.

model	dev	test	Iter
CharLSTM+LSTM+CRF	94.99%	94.38%	18/29
Elmo+LSTM+CRF	97.13%	96.37%	23/34
Parser+CharLSTM+LSTM+CRF	96.41%	96.48%	10/21
Parser+Elmo+LSTM+CRF	96.98%	96.59%	12/23
Bert+CharLSTM+LSTM+CRF
Bert+Elmo+CRF

Task:POS

Data:WSJ

Pretrained Embedding:glove.6B.100d.txt.

model	dev	test	Iter
CharLSTM+LSTM+CRF	97.67%	97.67%	10/21
Elmo+LSTM+CRF	97.89%	97.78%	3/14
Parser+ChatLSTM+LSTM+CRF	97.89%	97.70%	6/17
Parser+Elmo+LSTM+CRF	97.92%	97.78%	2/13

​