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 https://github.com/guillaumegenthial/sequence_tagging 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
The directory ./data_path contains:
- the preprocessed data files,
train_dataandtest_data - a vocabulary file
word2id.pklthat maps each character to a unique id
For generating vocabulary file, please refer to the code in data.py.
Each data file should be in the following format:
中 B-LOC
国 I-LOC
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
python main.py --mode=train
python main.py --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:
| P | R | F | F (PER) | F (LOC) | F (ORG) |
|---|---|---|---|---|---|
| 0.8945 | 0.8752 | 0.8847 | 0.8688 | 0.9118 | 0.8515 |
python main.py --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