Skip to content
A very simple BiLSTM-CRF model for Chinese Named Entity Recognition 中文命名实体识别 (TensorFlow)
Branch: master
Clone or download
Latest commit 4083fdf Mar 16, 2019
Type Name Latest commit message Commit time
Failed to load latest commit information.
.gitignore update Mar 15, 2018 update Mar 16, 2019 first commit Aug 22, 2017

A simple BiLSTM-CRF model for Chinese Named Entity Recognition

This repository includes the code for buliding a very simple character-based BiLSTM-CRF sequence labeling model for Chinese Named Entity Recognition task. Its goal is to recognize three types of Named Entity: PERSON, LOCATION and ORGANIZATION.

This code works on Python 3 & TensorFlow 1.2 and the following repository gives me much help.


This model is similar to the models provied by paper [1] and [2]. Its structure looks just like the following illustration:


For one Chinese sentence, each character in this sentence has / will have a tag which belongs to the set {O, B-PER, I-PER, B-LOC, I-LOC, B-ORG, I-ORG}.

The first layer, look-up layer, aims at transforming each character representation from one-hot vector into character embedding. In this code I initialize the embedding matrix randomly and I know it looks too simple. We could add some language knowledge later. For example, do tokenization and use pre-trained word-level embedding, then every character in one token could be initialized with this token's word embedding. In addition, we can get the character embedding by combining low-level features (please see paper[2]'s section 4.1 and paper[3]'s section 3.3 for more details).

The second layer, BiLSTM layer, can efficiently use both past and future input information and extract features automatically.

The third layer, CRF layer, labels the tag for each character in one sentence. If we use a Softmax layer for labeling, we might get ungrammatic tag sequences beacuse the Softmax layer labels each position independently. We know that 'I-LOC' cannot follow 'B-PER' but Softmax doesn't know. Compared to Softmax, a CRF layer can use sentence-level tag information and model the transition behavior of each two different tags.


#sentence #PER #LOC #ORG
train 46364 17615 36517 20571
test 4365 1973 2877 1331

It looks like a portion of MSRA corpus. I downloaded the dataset from the link in ./data_path/original/link.txt

data files

The directory ./data_path contains:

  • the preprocessed data files, train_data and test_data
  • a vocabulary file word2id.pkl that maps each character to a unique id

For generating vocabulary file, please refer to the code in

data format

Each data file should be in the following format:

很	O
大	O

句	O
子	O
结	O
束	O
是	O
空	O
行	O

If you want to use your own dataset, please:

  • transform your corpus to the above format
  • generate a new vocabulary file

How to Run


python --mode=train


python --mode=test --demo_model=1521112368

Please set the parameter --demo_model to the model that you want to test. 1521112368 is the model trained by me.

An official evaluation tool for computing metrics: here (click 'Instructions')

My test performance:

0.8945 0.8752 0.8847 0.8688 0.9118 0.8515


python --mode=demo --demo_model=1521112368

You can input one Chinese sentence and the model will return the recognition result:



[1] Bidirectional LSTM-CRF Models for Sequence Tagging

[2] Neural Architectures for Named Entity Recognition

[3] Character-Based LSTM-CRF with Radical-Level Features for Chinese Named Entity Recognition


You can’t perform that action at this time.
You signed in with another tab or window. Reload to refresh your session. You signed out in another tab or window. Reload to refresh your session.