Implementation of End-to-End Relation Extraction using LSTMs on Sequences and Tree Structures.

• Fedora Core 22
• clang++ 3.4
• boost 1.57
• yaml-cpp 0.5.1
• ICU4C 54.1

This code may work on other linux environments, but we have not tried them. For convenience, this package includes snapshot versions of clab/cnn (https://github.com/clab/cnn) and eigen (https://eigen.tuxfamily.org/). These follow the original license.

tar xzf cnn.tar.gz
tar xzf eigen.tar.gz
mkdir build
cd build
cmake .. -DEIGEN3_INCLUDE_DIR=eigen -DCMAKE_CXX_COMPILER=/usr/bin/clang++
make
cd ..

mkdir dict models

cd dict/
wget https://tti-coin.jp/data/wikipedia200.bin
cd ..

see data/README.md

These models are trained on the environments above (with Intel CPU). If you use other environments, please retrain the models to avoid the incompatibility problem.

cd models/
wget https://tti-coin.jp/data/ace2005-test.txt.gz
gunzip ace2005-test.txt.gz
cd ..

cd models/
wget https://tti-coin.jp/data/semeval-test.txt.gz
gunzip semeval-test.txt.gz
cd ..

Prediction results will be written as *.pred.ann in the test corpus directory.

build/relation/RelationExtraction --test -y yaml/parameter-ace2005.yaml

build/relation/RelationExtraction --test -y yaml/parameter-semeval-2010.yaml

build/relation/RelationExtraction --train -y yaml/parameter-ace2005.yaml

build/relation/RelationExtraction --train -y yaml/parameter-semeval-2010.yaml

YAML files for ACE2004 are not included. Please modify yaml/parameter-ace2005.yaml.

Scores may not be consistent with those in our paper due to the differences in the environments.

Please cite our ACL paper when using this software.

• Makoto Miwa and Mohit Bansal. End-to-end Relation Extraction using LSTMs on Sequences and Tree Structures. In the Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 1105--1116, 2016.