This repo contains code for the following paper:
Install the following Python 3 packages:
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
.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