Skip to content

Latest commit

 

History

History
48 lines (42 loc) · 3.43 KB

README.md

File metadata and controls

48 lines (42 loc) · 3.43 KB
Raw

Neural Tensor Network for Knowledge Base Completion with GRAKN.AI

Credits for the (bulk of the) neural network file code go to Siddarth Agarwal. His GitHub repo for this project is https://github.com/siddharth-agrawal/Neural-Tensor-Network

Credits for the neural tensor network project itself go to http://nlp.stanford.edu/~socherr/SocherChenManningNg_NIPS2013.pdf

To run this program, you will first need to have installed Grakn (this was written on 0.15.0). If you're coming here from the blog, then you probably already have grakn installed, but if you don't, you can use something like Homebrew to quickly get up and running, using the command

brew install grakn

You will then have to go into the Grakn directory and start the Grakn shell script, like so

/YOUR-GRAKN-DIRECTORY/bin/grakn.sh start

This will allow you to make queries through the Graql shell

Running the program

You will need Python 2. for this project. I cannot guarantee compatibility with Python 3.

First, navigate to the project directory and type pip install -r requirements.txt to install all the modules necessary to run this project

Next you must go into insertLexicon.py and at the top of the file, replace the PATH-TO-GRAKN in

_GRAQL_PATH = "/PATH-TO-GRAKN/bin/graql.sh"

with the directory path and name that you used to start the Grakn engine itself. For example, on my machine this line looks like

_GRAQL_PATH = "/Users/nickpowell/Documents/Grakn/bin/graql.sh"

You can run either the neuralTensorNetwork.py file, or the insertLexicon.py file. insertLexicon.py is a stand-alone separate from the neural netwrok that loads the ontology, inserts entities and relations, and checks for initial inferences. neuralTensorNetwork.py does the work of insertLexicon.py as well as that of the neural network. In most scenarios you'd want to run neuralTensorNetwork.py. The rest of this section assumes this is the file you are executing.

If it is your first time running the program, you will need to run it with the buildGraph flag on in order to build the ontology and ruleset. You will also need to specify a Graql keyspace with -k. The shell command looks like this:

python ./neuralTensorNetwork.py -k insert_your_keyspace_here --buildGraph

If you have multiple versions of Python installed, you may want to specify the 2. version, for example

python2.7 ./neuralTensorNetwork.py -k insert_your_keyspace_here --buildGraph

Once you have loaded the ontology and ruleset for a particular keyspace, REMOVE the --buildGraph flag from your command and on every subsequent program execution to that keyspace, simply use

python ./neuralTensorNetwork.py -k insert_your_keyspace_here

If you try to add the ontology to a keyspace that already has the ontology loaded, you may encounter errors, especially if you have made changes to the ontology.gql file.

Be patient when running this program. With 200 epochs, the NN takes ~6-8 hours to train - remember that this implementation does not make use TF or Theano. Each round of Grakn inference checks takes ~2 hours, depending on how many entities you are checking and how big your graph is. Each round of likely triplet searching takes about 1 hour as well. Inserting relations into the graph is much faster than these other tasks but is also non-trivial.