LSTM-ER

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

Requirements

• 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 (http://eigen.tuxfamily.org/). These follow the original license.

Usage

Compilation

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 ..

Creation of directories

mkdir dict models

Preparation of the pretrained embeddings

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

Preparation of the data sets

see data/README.md

Obtaining pretrained models

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.

ACE 2005 (Relation extraction)

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

SemEval 2010 Task 8 (Relation classification)

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

Testing models

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

ACE 2005 (Relation extraction)

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

SemEval 2010 Task 8 (Relation classification)

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

Training models

ACE 2005 (Relation extraction)

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

SemEval 2010 Task 8 (Relation classification)

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

Notes

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.