JointNRE

Codes and datasets for our paper "Neural Knowledge Acquisition via Mutual Attention between Knowledge Graph and Text"

Some Introduction

This implementation is a fast and stable version.

We have made some simplifications for the original model so that to train a joint model just needs around 15min.

We also encapsulate more neural architectures into our framework to encode sentences.

The code and datasets mainly for the task relation extraction.

Data

We provide the datasets used for the task relation extraction.

New York Times Corpus: The data used in relation extraction from text is published by "Modeling relations and their mentions without labeled text". The data should be obtained from [LDC] first.

Datasets are required in the folder data/ in the following format, containing at least 4 files:

kg/train.txt: the knowledge graph for training, format (e1, e2, rel).
text/relation2id.txt: the relation needed to be predicted for RE, format (rel, id).
text/train.txt: the text for training, format (e1, e2, name1, name2, rel, sentence).
text/vec.txt: the initial word embeddings.
[Download (Baidu Cloud)]
[Download (Tsinghua Cloud)]

For FB15K-NYT, we directly give the data for our code [Download (Tsinghua Cloud)], as we cannot release the original data limited by the license of LDC.

Run the experiments

To run the experiments, unpack the datasets first:

unzip origin_data.zip -d origin_data/
mkdir data/
python initial.py

Run the corresponding python scripts to train models:

cd jointE
bash make.sh
python train.py

Change the corresponding python code to set hyperparameters:

tf.app.flags.DEFINE_float('nbatch_kg',100,'entity numbers used each training time')
tf.app.flags.DEFINE_float('margin',1.0,'entity numbers used each training time')
tf.app.flags.DEFINE_float('learning_rate_kg',0.001,'learning rate for kg')
tf.app.flags.DEFINE_float('ent_total',lib.getEntityTotal(),'total of entities')
tf.app.flags.DEFINE_float('rel_total',lib.getRelationTotal(),'total of relations')
tf.app.flags.DEFINE_float('tri_total',lib.getTripleTotal(),'total of triples')
tf.app.flags.DEFINE_float('katt_flag', 1, '1 for katt, 0 for att')

tf.app.flags.DEFINE_string('model', 'cnn', 'neural models to encode sentences')
tf.app.flags.DEFINE_float('max_length',config['fixlen'],'maximum of number of words in one sentence')
tf.app.flags.DEFINE_float('pos_num', config['maxlen'] * 2 + 1,'number of position embedding vectors')
tf.app.flags.DEFINE_float('num_classes', config['textual_rel_total'],'maximum of relations')

tf.app.flags.DEFINE_float('hidden_size',230,'hidden feature size')
tf.app.flags.DEFINE_float('pos_size',5,'position embedding size')

tf.app.flags.DEFINE_float('max_epoch',20,'maximum of training epochs')
tf.app.flags.DEFINE_float('batch_size',160,'entity numbers used each training time')
tf.app.flags.DEFINE_float('learning_rate',0.5,'learning rate for nn')
tf.app.flags.DEFINE_float('weight_decay',0.00001,'weight_decay')
tf.app.flags.DEFINE_float('keep_prob',0.5,'dropout rate')

tf.app.flags.DEFINE_string('model_dir','./model/','path to store model')
tf.app.flags.DEFINE_string('summary_dir','./summary','path to store summary_dir')

Run the corresponding python scripts to test models:

cd jointE
bash make.sh
python test.py

Note that the hyperparameters in the train.py and the test.py must be the same.

Run the corresponding python script to get PR-curve results:

cd jointE
python pr_plot.py

Citation

 @inproceedings{han2018neural,
   title={Neural Knowledge Acquisition via Mutual Attention between Knowledge Graph and Text},
   author={Han, Xu and Liu, Zhiyuan and Sun, Maosong},
   booktitle={Proceedings of AAAI},
   year={2018}
 }

zc-young/JointNRE