Tutorial: Getting Started

Once you’ve built and installed Divisi, you can start by making your own AnalogySpace.

1. Download the ConceptNet tensor for English.

2. Load it:

   from csc.util.persist import get_picklecached_thing
   tensor = get_picklecached_thing('tensor.gz')
   

3. Run the SVD:

   svd = tensor.svd(k=100)
   

4. Get similar concepts (to ‘teach’):

   svd.u_dotproducts_with(svd.weighted_u_vec('teach')).top_items(10)
   

5. Predict properties (for ‘trumpet’):

   svd.v_dotproducts_with(svd.weighted_u_vec('trumpet')).top_items(10)
   

(These things that look like ('right', 'IsA', 'pet') are how we represent the features in ConceptNet.)

6. Evaluate possible assertions:

   # Is a dog a pet?
   svd.get_ahat(('dog', ('right', 'IsA', 'pet')))
   
   # Is a hammer a pet?
   svd.get_ahat(('hammer', ('right', 'IsA', 'pet')))
   

Note

The concepts are represented by their normalized form, which is how they are represented internally in ConceptNet.

If you look up a value for “scissors”, for example, you won’t find anything. It’s under “scissor”. The reason for this is so that phrases like “eat a sandwich” and “eating sandwiches” correspond to the same concept. See Natural language tools for details on how this works.