taominer / inferwiki Goto Github PK

Hi @TaoMiner, I appreciate very much your work and especially your publication related to inferwiki. I have added the knowledge base InferWiki16k and InferWiki64k to the mkb library that I created for my studies on KGE and Language Models. The objective of this tool is to facilitate model evaluation doing link prediction and triplet classification.

Both datasets are available on lastest dev version:

pip install git+https://github.com/raphaelsty/mkb --upgrade

>>> from mkb import datasets

>>> dataset = datasets.InferWiki16k(batch_size = 256)

>>> dataset
InferWiki16k dataset
        Batch size  256   
          Entities  15472 
         Relations  197   
           Shuffle  True  
     Train triples  162424
Validation triples  3398  
      Test triples  3398  

>>> from mkb import datasets

>>> dataset = datasets.InferWiki64k(batch_size = 256)

>>> dataset
InferWiki64k dataset
        Batch size  256   
          Entities  61944 
         Relations  239   
           Shuffle  True  
     Train triples  782443
Validation triples  7747  
      Test triples  7747  

Maybe this will be useful to some of your readers, have a nice day,

Raphaël

I noticed a minor inconsistency between the 16k and 64k datasets.

The InferWiki64k documentation says "InferWiki64k has the same format with InferWiki16k..." however for test_pos.txt and valid_pos.txt the relations have a path length element in 16k, but not in 64k.

I'm not sure if this is a mistake in the dataset and 64k should have path lengths too or if the mistake is in the documentation.