Joe Henke's 6.UAP Project

A tool to visualize inference over semantic networks such as ConceptNet.

Powered by Celestrium.

Setup

Prerequisites

Install virtualenv and the Heroku Toolbelt.

Installation

# clone this repo and its submodules
git clone --recursive https://github.com/jdhenke/uap.git && cd uap

# install node packages
npm install

# setup virtual environment
virtualenv env
source env/bin/activate

# install normal dependencies
pip install -r requirements.txt

# install these *special* dependencies
# this is a result of them depending on numpy in their setup.py's
pip install divisi2 csc-pysparse

Interface Explanation

• Add nodes using the search box in the top right.
• Select/Deselect nodes by clicking them and using the hotkeys Ctrl A and Esc
• Add related nodes by selecting nodes and using the hotkey + which is really Shift =.
• Remove nodes by selecting them and using the hotkey Delete.

Play with the sliders in the top left to get a feel for what they do.

You can also drag the white bar on the link strength distribution chart.

Visualizing Conceptnet

Run the following from the root of this repo:

foreman start

Then go to http://localhost:5000 to visualize assertions found in Conceptnet.

You should do this to compile the coffeescript files used in this project. See server.sh to see how to do this manually.

Visualizing Your Knowledgebase

Creating the Pickled Sparse Matrix

You will need to create a pickled sparse matrix representing your knowledgebase.

To easily do this, modify getAssertions() in src/parse.py to return a list or be a generator of assertions in your knowledgebase.

It works out of the box with a very simple knowledgebase if you want to try using that first to get a feel for the process.

Then run the following from the root of this repo to create ./kb.pickle.

python src/parse.py

Using the Pickled Sparse Matrix

Now to visualize the concepts in your knowledgebase based on sampling results at dimensionalities of 1, 2 and 3, run the following from the root of this repo:

python src/server.py ./kb.pickle 1,2,3 concepts src/www 5000

Again, go to http://localhost:5000/.

Here's an explanation of the command line arguments:

• ./kb.pickle is the path to the pickled sparse matrix file of your knowledgebase
• 25,50,100 is the comma delimited list of dimensions at which to sample the inference results.
• concepts is the node type, which indicates what nodes will be in the graph and must be either concepts or assertions.
• src/www is the path to the directory that will be served statically - should be where your index.html is.
• 5000 is the port that the visualization will be served on

Debugging

If you can't click in the node search box, try clicking the middle of the window then clicking in the box. Seems like a tyepahead issue.

If your interface isn't changing between concepts and assetions when you launch your server differently, you probably need to clear your browser's cache.

Recommended Reading

• License
• Deploying to Heroku