Awesome

NER-BERT-pytorch

PyTorch solution of Named Entity Recognition task with Google AI's BERT model.

利用Google AI的BERT模型进行中文命名实体识别任务的PyTorch实现。

Welcome to watch, star or fork.

MSRA dataset

Here, we take the Chinese NER data MSRA as an example. Of course, the English NER data is also fully applicable.

Named entity recognition task is one of the tasks of the Third SIGHAN Chinese Language Processing Bakeoff, we take the simplified Chinese version of the Microsoft NER dataset as the research object.

Data Formats

The NER dataset of MSRA consists of training set data/msra_train_bio and test set data/msra_test_bio, and no validation set is provided. There are 45000 training samples and 3442 test samples, and we will divide them appropriately later.

The dataset contains three types of entities: Person, Organization, Location and Other, the corresponding abbreviated tags are PER, ORG and LOC and O.

The format is similar to that of the Co-NLL NER task 2002, adapted for Chinese. The data is presented in two-column format, where the first column consists of the character and the second is a tag. The tag is specified as follows:

Tag	Meaning
O	Not part of a named entity
B-PER	Beginning character of a person name
I-PER	Non-beginning character of a person name
B-ORG	Beginning character of an organization name
I-ORG	Non-beginning character of an organization name
B-LOC	Beginning character of a location name
I-LOC	Non-beginning character of a location name
B-GPE	Beginning character of a geopolitical entity
I-GPE	Non-beginning character of a geopolitical entity

Dataset patition

We randomly select 3000 samples from the training set as the validation set, and the test set is unchanged. Thus, the dataset distribution is as follows.

Dataset	Number
training set	42000
validation set	3000
test set	3442

Requirements

This repo was tested on Python 3.5+ and PyTorch 0.4.1/1.0.0. The requirements are:

tensorflow >= 1.11.0
torch >= 0.4.1
pytorch-pretrained-bert == 0.4.0
tqdm
apex

Note: The tensorflow library is only used for the conversion of pre-trained models from TensorFlow to PyTorch. apex is a tool for easy mixed precision and distributed training in Pytorch, please see https://github.com/NVIDIA/apex.

Results

We didn't search best parameters and obtained the following results.

Overall results

Based on the best performance of the model on the validation set, the overall effect of the model is as follows:

Dataset	F1_score
training set	99.88
validation set	95.90
test set	94.62

Detail results on test set

Based on the best model on the validation set, we can get the recognition effect of each entity type on the test set.

NE Types	Precison	Recall	F1_score
PER	96.36	96.43	96.39
ORG	89.64	92.07	90.84
LOC	95.92	95.13	95.52

Usage

Get BERT model for PyTorch

There are two ways to get the pre-trained BERT model in a PyTorch dump for your experiments :
- Direct download of the converted pytorch version of the BERT model
  
  You can download the pytorch dump I converted from the tensorflow checkpont from my Google Cloud Drive folder bert-base-chinese-pytorch, including the BERT parameters file bert_config.json, the model file pytorch_model.bin and the vocabulary file vocab.txt.
- Convert the TensorFlow checkpoint to a PyTorch dump by yourself
  - Download the Google's BERT base model for Chinese from BERT-Base, Chinese (Chinese Simplified and Traditional, 12-layer, 768-hidden, 12-heads, 110M parameters), and decompress it.
  - Execute the following command, convert the TensorFlow checkpoint to a PyTorch dump.
```
export TF_BERT_BASE_DIR=/path/to/chinese_L-12_H-768_A-12
export PT_BERT_BASE_DIR=/path/to/NER-BERT-pytorch/bert-base-chinese-pytorch

pytorch_pretrained_bert convert_tf_checkpoint_to_pytorch \
	$TF_BERT_BASE_DIR/bert_model.ckpt \
	$TF_BERT_BASE_DIR/bert_config.json \
	$PT_BERT_BASE_DIR/pytorch_model.bin
```
  - Copy the BERT parameters file bert_config.json and dictionary file vocab.txt to the directory $PT_BERT_BASE_DIR.
```
cp $TF_BERT_BASE_DIR/bert_config.json $PT_BERT_BASE_DIR/bert_config.json
cp $TF_BERT_BASE_DIR/vocab.txt $PT_BERT_BASE_DIR/vocab.txt
```
Build dataset and tags
```
python build_msra_dataset_tags.py
```
It will extract the sentences and tags from the dataset data/msra_train_bio and data/msra_test_bio, split them into train/val/test and save them in a convenient format for our model, and create a file tags.txt containing a collection of tags.
Set experimental hyperparameters

We created a base_model directory for you under the experiments directory. It contains a file params.json which sets the hyperparameters for the experiment. It looks like
```
{
    "full_finetuning": true,
    "max_len": 180,

    "learning_rate": 3e-5,
    "weight_decay": 0.01,
    "clip_grad": 5,
}
```
For every new experiment, you will need to create a new directory under experiments with a params.json file.
Train and evaluate your experiment

if you use default parameters, just run
```
python train.py
```
Or specify parameters on the command line
```
python train.py --data_dir data/msra --bert_model_dir bert-base-chinese-pytorch --model_dir experiments/base_model --multi_gpu
```
It will instantiate a model and train it on the training set following the hyperparameters specified in params.json. It will also evaluate some metrics on the development set.
Evaluation on the test set

Once you've run many experiments and selected your best model and hyperparameters based on the performance on the development set, you can finally evaluate the performance of your model on the test set.

if you use default parameters, just run
```
python evaluate.py
```
Or specify parameters on the command line
```
python evaluate.py --data_dir data/msra --bert_model_dir bert-base-chinese-pytorch --model_dir experiments/base_model
```

References

Devlin et al. BERT: Pre-training of Deep Bidirectional Trasnsformers for Language Understanding (2018) [paper]
google-research/bert [github]
huggingface/pytorch-pretrained-BERT [github]
NVIDIA/apex [github]
chakki-works/seqeval [github]