BERT-NER

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

Requirements

• python 3.5+
• pytorch 0.4.1
• pytorch-pretrained-bert 0.4.0
• tqdm
• PyYAML

Overview

• njuner
  • An Chinese 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.

NJUNER

An Chinese NERer which recognizes PERSONs, LOCATIONs and ORGANIZATIONs in texts. It is completely character-based and does not require word segmentation or part-of-speech information.

Installation

pip install njuner

Usage

• 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 the model pretrained on the MSRA NER dataset from the NJUNER releases page. Uncompress the model archive and pass the directory to the parameter "model_dir".

Performance

Metrics: Span-based F1

• Training and testing on the corresponding dataset.

  Item	MSRA	Weibo-NE
  SOTA	93.18	55.28
  NJUNER	94.78	66.95

  • Results of SOTA are according to the paper Chinese NER Using Lattice LSTM.
  • Our model fined tune on the BERT, which pretrained on large-scale unlabeled corpus, so the above results are not strictly comparable.

• 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".