OpenNRE

An open-source framework for neural relation extraction.

Contributed by Tianyu Gao, Xu Han, Lumin Tang, Yankai Lin, Zhiyuan Liu

Overview

It is a TensorFlow-based framwork for easily building relation extraction models. We divide the pipeline of relation extraction into four parts, which are embedding, encoder, selector and classifier. For each part we have implemented several methods.

Embedding
- Word embedding
- Position embedding
- Concatenation method
Encoder
- PCNN
- CNN
- RNN
- BiRNN
Selector
- Attention
- Maximum
- Average
Classifier
- Softmax loss function
- Output

All those methods could be combined freely.

We also provide fast training and testing codes. You could change hyper-parameters or appoint model architectures by using Python arguments. A plotting method is also in the package.

Advesarial training method is implemented following Wu et al. (2017). Since it's a general training method, you could adapt it to any models with simply adding a few lines of code.

This project is under MIT license.

Requirements

Python (>=2.7)
TensorFlow (>=1.4.1)
- CUDA (>=8.0) if you are using gpu
Matplotlib (>=2.0.0)
scikit-learn (>=0.18)

Installation

Install TensorFlow
Clone the OpenNRE repository:

git clone git@github.com:thunlp/OpenNRE.git

Download NYT dataset from Google Drive or Tsinghua Cloud
Extract dataset to ./origin_data

tar xvf origin_data.tar

Results

F1 Score Results

Encoder\Selector(Trainer)	Attention	Attention(Adv)	Maximum	Average
PCNN	0.452	0.456	0.443	0.439
CNN	0.431	0.445	0.430	0.438
RNN	0.439	0.453	0.436	0.445
BiRNN	0.427	0.447	0.438	0.442

(Adv) means using adversarial training

AUC Results

Encoder\Selector(Trainer)	Attention	Attention(Adv)	Maximum	Average
PCNN	0.408	0.416	0.406	0.392
CNN	0.381	0.392	0.386	0.383
RNN	0.385	0.402	0.380	0.408
BiRNN	0.367	0.389	0.368	0.388

Quick Start

Process Data

python gen_data.py

The processed data will be stored in ./data

HINT: If you are using python3, execute python3 gen_data_python3.py.

Train Model

python train.py --model_name pcnn_att

The arg model_name appoints model architecture, and pcnn_att is the name of one of our models. All available models are in ./model. About other arguments please refer to ./train.py. Once you start training, all checkpoints are stored in ./checkpoint.

Test Model

python test.py --model_name pcnn_att

Same usage as training. When finishing testing, the best checkpoint's corresponding pr-curve data will be stored in ./test_result.

Plot

python draw_plot.py pcnn_att

The plot will be saved as ./test_result/pr_curve.png. You could appoint several models in the arguments, like python draw_plot.py pcnn_att pcnn_max pcnn_ave, as long as there are these models' results in ./test_result.

Build Your Own Model

Not only could you train and test existing models in our package, you could also build your own model or add methods to the four basic modules. When adding a new model, you could create a python file in ./model having the same name as the model and implement it like following:

from framework import Framework
import tensorflow as tf

def your_new_model(is_training):
    if is_training:
        framework = Framework(is_training=True)
    else:
        framework = Framework(is_training=False)

    # Word Embedding
    word_embedding = framework.embedding.word_embedding()
    # Position Embedding. Set simple_pos=True to use simple pos embedding
    pos_embedding = framework.embedding.pos_embedding()
    # Concat two embeddings
    embedding = framework.embedding.concat_embedding(word_embedding, pos_embedding)
    
    # PCNN. Appoint activation to whatever activation function you want to use.
    # There are three more encoders:
    #     framework.encoder.cnn
    #     framework.encoder.rnn
    #     framework.encoder.birnn
    x = framework.encoder.pcnn(embedding, FLAGS.hidden_size, framework.mask, activation=tf.nn.relu)
    
    # Selective attention. Setting parameter dropout_before=True means using dropout before attention. 
    # There are three more selecting method
    #     framework.selector.maximum
    #     framework.selector.average
    #     framework.selector.no_bag
    logit, repre = framework.selector.attention(x, framework.scope, framework.label_for_select)

    if is_training:
        loss = framework.classifier.softmax_cross_entropy(logit)
        output = framework.classifier.output(logit)
        # Set optimizer to whatever optimizer you want to use
        framework.init_train_model(loss, output, optimizer=tf.train.GradientDescentOptimizer)
        framework.load_train_data()
        framework.train()
    else:
        framework.init_test_model(tf.nn.softmax(logit))
        framework.load_test_data()
        framework.test()

After creating model's python file, you need to add the model to ./train.py and ./test.py as following:

# other code ...

def main():
    from model.your_new_model import your_new_model

# other code ...

Then you can train, test and plot!

As for using adversarial training, please refer to ./model/pcnn_att_adv.py for more details.

Reference

Neural Relation Extraction with Selective Attention over Instances. Yankai Lin, Shiqi Shen, Zhiyuan Liu, Huanbo Luan, Maosong Sun. ACL2016. paper
Adversarial Training for Relation Extraction. Yi Wu, David Bamman, Stuart Russell. EMNLP2017. paper
A Soft-label Method for Noise-tolerant Distantly Supervised Relation Extraction. Tianyu Liu, Kexiang Wang, Baobao Chang, Zhifang Sui. EMNLP2017. paper
Reinforcement Learning for Relation Classification from Noisy Data. Jun Feng, Minlie Huang, Li Zhao, Yang Yang, Xiaoyan Zhu. AAAI2018. paper

hankers / OpenNRE