Directly extract relation triples from raw text. Running on Linux.

We prepared a demo website. You can input some sentences to get relation triples.

python triple.py

or

import triple
triple.triple("Mr. Scheider played the police chief of a resort town menaced by a shark.")
# output: triples: [['Scheider', 'per:title', 'police chief', 11.785091400146484]]

Linux Shell

Java 1.8+ (Check with command: java -version) (Download Page)

Python 3

CUDA >= 9.0 (Check with command: nvcc --version)(Download Page)

Clone the repository from OpenNRE github page:

git clone https://github.com/thunlp/OpenNRE.git --depth 1

Copy modified frame into OpenNRE:

cp sentence_re.py OpenNRE/opennre/framework/sentence_re.py

Then install OpenNRE:

cd OpenNRE
pip install -r requirements.txt
python setup.py install 
cd ..

Download Pretrained file:

cd pretrain
bash download_bert.sh
cd ..

Install using pip:

pip install stanfordcorenlp

Download and upzip Stanfordcore:

wget -c http://nlp.stanford.edu/software/stanford-corenlp-full-2018-02-27.zip
upzip stanford-corenlp-full-2018-02-27.zip
rm stanford-corenlp-full-2018-02-27.zip

Then install NLTK:

pip install nltk

Download tokenizer:

python
import nltk
nltk.download('punkt')
exit()

Due to copyright of the Tacred data, we are not able to share the data or our distant supervision based crawled data with you, please download it by yourself.

Get Stanford Tacred Data(https://nlp.stanford.edu/projects/tacred/) and put train.json, dev.json, test.josn under benchmark/tacred, then processing data:

cd ./benchmark/tacred
python data.py
cd ../../

If you want to crawl distant supervised data by Google search engine, please do as following steps:

1). Use Google Search API to crawl articles based on distant supervision. Please modify the api key and filename in google_crawling.py.

python ./Google_Crawler/google_crawling.py

2).Split the target sentences from the articles. Please modify the filename in processing.py.

python ./Google_Crawler/processing.py

If you want to train your own model, follow this step.

Run python file to start training:

python train_tacred_bert_softmax.py

Please modify batch size in line 44 in order to match the graphic memory on your machine.

Download pretrained model file:

mkdir ckpt
cd ckpt
wget -c https://www.dropbox.com/s/7f70dy2vatmmly4/tacred_bert_softmax.pth.tar
cd ..

Then, simply run

python triple.py

to get relation triples from raw sentence.

To run on a server:

python server.py --port 12345

Change the port and send GET request to server to get result like:

localhost:12345?content=Mr. Scheider played the police chief of a resort town menaced by a shark.