Skip to content

Latest commit

 

History

History
91 lines (62 loc) · 3.55 KB

README.md

File metadata and controls

91 lines (62 loc) · 3.55 KB
Raw

Aqqu Question Answering System

This is an ancestor of the code accompanying the publication: More Accurate Question Answering on Freebase, Hannah Bast and Elmar Haussmann, CIKM 2015

The original version can be found as as the very first commit of this repository and under the v1.0 tag.

Follow the instructions below to set up the system. This also includes descriptions on how to obtain pre-requisite data.

Setup is easy if all pre-requisites are met.

Requirements:

  • Docker > 18.09 on 64 bit Linux
  • RAM: 40 GB for training the large WebQuestions/WebQSP models
  • Disk: about 50 GB for all required data

Setup a Virtuoso Instance with Freebase

To setup a Virtuoso instance with the Freebase data needed for Aqqu we recommend to follow the instructions at our Virtuoso with Docker Compose repository. This will automatically download the correct version of Freebase and setup Virtuoso with the exact same settings used by us.

Get the Dataset

All data required for learning can be found under /nfs/datastets/aqqu_input_data when on any of the Chair's computer systems, all other data is generated automatically.

cp -r /nfs/datasets/aqqu_input_data/* input/

Outside our system's please contact Prof. Hannah Bast. with the above path to get a download.

Train with the provided script

When using docker/wharfer with user namespaces you may need to first run chmod o+wX data so Aqqu running inside docker can write data even if it is a very restricted user like nobody

./build_and_run.sh train -n <user_provided_name> -r <ranker e.g. WQSP_Ranker> <additional args>

Run with the provided script

./build_and_run.sh backend -n <user_provided_name> -r <ranker e.g. WQSP_Ranker> -p <port> <addtional args>

Debug changes with the provided script

The backend command does not rebuild the image but reuses the exact image used for training. Therefore changes to the source code made after the train step are not reflected in the behavior of backend.

To try out changes without affecting the train/backend image use the debug command.

./build_and_run.sh debug -n <user_provided_name> -r <ranker e.g. WQSP_Ranker> -p <port> <addtional args>

Once the changes have been tested the image used by the backend command can be updated using. Note however that this is somewhat dangerous as it does not work for changes that would make the model incompatible to the trained model.

./build_and_run.sh update -n <user_provided_name>

Test the Backend

The Aqqu backend provides a simple JSON API that can easily be tested using curl. If you have the ./build_and_run.sh backend … command running on a server <host> with port <port> the following asks for Albert Einstein's place of birth

curl http://<host>:<port>/?q=where%20was%20albert%einstein%20born

Run Cross Validation with the provided script

./build_and_run.sh cv -n <user_provided_name> -r <ranker> <dataset name>

Overriding parameters

To override certain ranker parameters you can use --override with a JSON object as additional argument for example --override '{"top_ngram_percentile": 15}'

Disabling GPU

To disable GPU use run above commands with the environment variable NO_GPU=1

Using (nvidia-)docker directly

To use (nvidia-)docker directly refer to the build_and_run.sh script. No additional documentation is provided as this is discouraged the script should be reasonably easy to follow.