An attempt to do better on Chinese NER task.

• Neural Architectures for Named Entity Recognition

  • with pretrained glove embedding
  • Bi-LSTM (character-level input)
  • (optional) LSTM (capture the features of Capitalization)
  • concatenate above two (or three) to form the real input
  • Bi-LSTM (word-level input and character-level input)
  • CRF loss layer

• Character-Based LSTM-CRF with Radical-Level Features for Chinese Named Entity Recognition

  • with pretrained word2vec(character2vec may be more proper) embedding
  • Bi-LSTM (radical-level input)
  • concatenate above two
  • Bi-LSTM (real input)
  • CRF loss layer

• A neural network model for Chinese named entity recognition
  See also in Improving Named Entity Recognition for Chinese Social Media with Word Segmentation Representation Learning

  • with pretrained word2vec(same as above one) embedding
  • do a segmentation operation to yield input
  • concatenate above two
  • Bi-LSTM (real input)
  • CRF loss layer

• NER-pytorch

  • a pytorch version implementation which is similar to the very beginning one.

• SeqGAN: Sequence Generative Adversarial Nets with Policy Gradient

  • a framework using GAN + RL to solve sequence problem.

• Adversarial Learning for Chinese NER from Crowd Annotations

  • 2018 AAAI Alibaba

• RL-GAN For NLP: 强化学习在生成对抗网络文本生成中扮演的角色

• Tensorflow
• pynlpir
• jieba

Note: Current repo only contains Version 1 which is discussed in Experiments, but you could easily modify the cofigurations in main.py to have the Baseline set in Experiments

1. Preprocess the dataset, whose format should be the same as the examples in /data. And the code in preprocess.py shall be considered as the refernce.
2. Yield a character-level embedding with Glove or fastText. Refer to the README.md in /Glove and /fastText to gain an intuition.
3. Modify configurations in main.py to get a brand new config suiting your idea.
4. Train your own model by

python main.py --train=True

5. Evaluate the outcome of your model by

python main.py

If you have any question about this repo or you find some problems in my code, feel free to contact with me via email or just give comments in Issues.

• Corpus: MSRA

• Tag schema: iobes

  我 S-PER
  来 O
  自 O
  北 B-ORG
  京 I-ORG
  理 I-ORG
  工 I-ORG
  大 I-ORG
  学 E-ORG

• Dataset split

  90% train, 10% test

• Structure

  • with pretrained GloVe embedding for characters and random initialized embedding for segmentations
  • concatenate above two
  • Bi-LSTM
  • CRF loss layer

• Result(Test only)

  Type	Accuracy	Precision	Recall	FB1
  LOC	\	94.11%	90.58%	92.31
  ORG	\	86.86%	88.43%	87.64
  PER	\	92.21%	92.45%	92.33
  OVER ALL	98.79%	91.89%	90.71%	91.30

• Corpus: SAME as baseline

• Tag schema: iobes

  我 S-PER
  来自 O
  北京 B-ORG
  理工 I-ORG
  大学 E-ORG

• Dataset split: SAME as baseline

• Structure

  • with pretrained GloVe embedding for words
  • use a layer of CNN to capture the features of characters projected by embedding
  • Bi-LSTM
  • CRF loss layer

• Result(Test only)

  Type	Accuracy	Precision	Recall	FB1
  LOC	\	93.31%	91.14%	92.21
  ORG	\	87.15%	86.63%	86.89
  PER	\	93.94%	91.94%	92.93
  OVER ALL	98.83%	92.17%	90.42%	91.29