Context-Aware Representations for Knowledge Base Relation Extraction

Relation extraction on an open-domain knowledge base

Accompanying repository for our EMNLP 2017 paper (full paper). It contains the code to replicate the experiments and the pre-trained models for sentence-level relation extraction.

This repository contains experimental software and is published for the sole purpose of giving additional background details on the respective publication.

Please use the following citation:

@inproceedings{TUD-CS-2017-0119,
	title = {Context-Aware Representations for Knowledge Base Relation Extraction},
	author = {Sorokin, Daniil and Gurevych, Iryna},
	booktitle = {Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
	pages = {(to appear)},
	year = {2017},
	location = {Copenhagen, Denmark},
}

Paper abstract:

We demonstrate that for sentence-level relation extraction it is beneficial to consider other relations in the sentential context while predicting the target relation. Our architecture uses an LSTM-based encoder to jointly learn representations for all relations in a single sentence. We combine the context representations with an attention mechanism to make the final prediction. We use the Wikidata knowledge base to construct a dataset of multiple relations per sentence and to evaluate our approach. Compared to a baseline system, our method results in an average error reduction of 24% on a held-out set of relations.

Please, refer to the paper for more details.

The dataset described in the paper can be found here:

• https://www.ukp.tu-darmstadt.de/data/lexical-resources/wikipedia-wikidata-relations/

Contacts:

If you have any questions regarding the code, please, don't hesitate to contact the authors or report an issue.

• Daniil Sorokin, lastname@ukp.informatik.tu-darmstadt.de
• https://www.ukp.tu-darmstadt.de
• https://www.tu-darmstadt.de

Demo:

You can try out the model on single sentences in our demo:

http://semanticparsing.ukp.informatik.tu-darmstadt.de:5000/relation-extraction/

Project structure:

relation_extraction/
├── eval.py
├── model-train-and-test.py
├── notebooks
├── optimization_space.py
├── core
│   ├── parser.py
│   ├── embeddings.py
│   ├── entity_extraction.py
│   └── keras_models.py
├── relextserver
│   └── server.py
├── graph
│   ├── graph_utils.py
│   ├── io.py
│   └── vis_utils.py
├── stanford_tag_dataset.py
└── evaluation
    └── metrics.py
resources/
├── properties-with-labels.txt
└── property_blacklist.txt

File	Description
relation_extraction/	Main Python module
relation_extraction/core	Models for joint relation extraction
relation_extraction/relextserver	The code for the web demo.
relation_extraction/graph	IO and processing for relation graphs
relation_extraction/evaluation	Evaluation metrics
resources/	Necessary resources
data/curves/	The precision-recall curves for each model on the held out data

Setup:

1. We recommend that you setup a new pip environment first: http://docs.python-guide.org/en/latest/dev/virtualenvs/

2. Check out the repository and run:

pip3 install -r requirements.txt

3. Set the Keras (deep learning library) backend to TensorFlow with the following command:

export KERAS_BACKEND=tensorflow

You can also permanently change Keras backend (read more: https://keras.io/backend/). Note that in order to reproduce the experiments in the paper you have to use Theano as a backend instead.

4. Download the data, if you want to replicate the experiments from the paper. Extract the archive inside emnlp2017-relation-extraction/data/wikipedia-wikidata/.

5. Download the GloVe embeddings, glove.6B.zip and put them into the folder emnlp2017-relation-extraction/resources/glove/.

Pre-trained models:

• You can download the models that were used in the experiments here
• See Using pre-trained models.ipynb for a detailed example on how to use the pre-trained models in your code

Reproducing the experiments from the paper

To reproduce the experiments please refer to the version of the code that was published with the paper: tag emnlp17

1. Complete the setup above

2. Run python model-train-and-test.py in emnlp2017-relation-extraction/relation_extraction/ to see the list of parameters

3. If you put the data into teh default folders you can train the ContextWeighted model with the following command:

python model-train-and-test.py model_ContextWeighted --mode train-test

Notes

• The web demo code is provided for information only. It is not meant to be run elsewhere.

Requirements:

• Python 3.4
• See requirements.txt for library requirements.

License:

• Apache License Version 2.0