This is a named entity recognizer based on pytorch-pretrained-bert.

• python 3.5+
• pytorch 0.4.1
• pytorch-pretrained-bert 0.6.1
• tqdm
• PyYAML

• njuner
  • A NER package, more details in section NJUNER.
• preprocess_msra.py
  • A tool for preprocessing the MSRA NER dataset.
• preprocess_pd98.py
  • A tool for preprocessing the People's Daily 1998.1 dataset.
• run_ner.py
  • A tool for training and testing bert-ner model on different datasets.
  • Working with task_config.yaml.

A NER tool which recognizes predefined entities like PERSONs, LOCATIONs and ORGANIZATIONs in texts. It is completely end-2-end and does not require word segmentation or part-of-speech information.

pip install njuner

• As a package

  from njuner import NJUNER
  ner = NJUNER(model_dir=model_path)
  ner.label(['李雷和韩梅梅去上海迪斯尼乐园。'])
  # [[('B-PER', '李'), ('I-PER', '雷'), ('O', '和'), ('B-PER', '韩'), ('I-PER', '梅'), ('I-PER', '梅'), ('O', '去'), ('B-ORG', '上'), ('I-ORG', '海'), ('I-ORG', '迪'), ('I-ORG', '斯'), ('I-ORG', '尼'), ('I-ORG', '乐'), ('I-ORG', '园'), ('O', '。') ]]

• As a command line tool

  • Manual

    njuner -h

  • An example

    njuner --model_dir model_path --input_file input.txt --output_dir ./

    This will produce there files, which are "tokens.txt", "predictions.txt" and "summary.txt", in the output directory.

• Pretrained model

  You can get several pretrained models from the NJUNER releases page. Uncompress the model archive and pass the directory to the parameter "model_dir".

• Training and testing on the corresponding dataset.

  Item	MSRA	Weibo-NE	Resume	CoNLL-2003
  Baseline	93.18	55.28	94.46	92.4
  NJUNER	95.02	66.95	95.62	91.7

  • Baselines of MSRA, Weibo-NE and Resume are reported in the paper Chinese NER Using Lattice LSTM, which are SOTA.
  • The baseline of CoNLL-2003 is reported in the original paper BERT, which we find is hard to reproduce.

• Comparison of different Chinese NER tools.

  Item	MSRA	Weibo-NE
  HanLP	72.65	38.66
  LTP	73.34	43.97
  NJUNER	81.58	63.08

  • The NER modules of HanLP and LTP are both trained on the People's Daily 1998.1 dataset. For comparison, NJUNER is also trained on the same dataset and achieve the above results on MSRA and Weibo-NE respectively. Also, their target entity types are same with our tool's, which are "PER", "LOC" and "ORG".
  • There is another entity type "GPE" in Weibo-NE dataset. For comparison, we uniformly refer to "GPE" as "LOC".