memozhu / buboqa

Simple question answering over knowledge graphs (Mohammed et al., NAACL 2018)

Simple Question Answering over Knowledge Graphs

This repo contains code for the following paper:

• Salman Mohammed, Peng Shi, and Jimmy Lin. Strong Baselines for Simple Question Answering over Knowledge Graphs with and without Neural Networks. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 291-296, May 2018, New Orleans, Louisiana.

Running the Code

Install the following Python 3 packages:

• PyTorch (version 0.4.0)
• torchtext (version 0.2.3)
• NLTK
• NLTK data (tokenizers, stopwords list)
• fuzzywuzzy

If you use PyTorch version 0.2.0, please checkout to

commit 34a71f29192ed57f83d8576002f2b540de7d722f

Run the setup script. This takes a long time. It fetches dataset, other files, processes them and creates indexes:

sh setup.sh 

There are four main components to our formulation of the problem, as detailed in the paper: entity detection, entity linking, relation prediction and evidence integration. Each of these components is contained in a separate directory, with an associated README.

• entity_detection and relation_prediction can be run independently.
• entity_detection needs to be run before entity_linking.
• entity_linking and relation_prediction needs to be run before evidence_integration.

Running the Code with Docker (GPU, Ubuntu 16, Cuda 9.0 base)

• Make sure you have the Docker daemon running

• Build the image from Dockerfile

cp docker_files/Dockerfile_gpu Dockerfile
docker build -t buboqa .

• Run the Docker image on GPU with nvidia-docker installed. Notice that we are mounting the current directory so that data persists.

nvidia-docker run -it --rm \
  -v "$(pwd)":/code \
  buboqa

• OR ... Run the Docker image on CPU (not tested)

docker run -it --rm \
  -v "$(pwd)":/code \
  buboqa

• Exit shell when needed

$  exit