zafarali / embedkb Goto Github PK

• Currently it takes too long to do the entity prediction task. This is because I score a large batch in tensorflow and then use numpy to argsort and stuff. Ideally the model itself should use tf.nn.top_k to filter out the top instances in tensorflow itself.

• Model saving should also save the embedding dimensions and stuff so that it can be quickly reloaded.

• NTN
• Unstructured embedding
• Structured embedding
• TransE
• TransH
• TransR
• TransD

Test to see if the implemented models can receive accuracy on the triple classification task as reported elsewhere.

TransE (Bordes et al. 2013) achieves 87.6% accuracy on FBK13 (as reported by Nguyen et al. 2017). We should achieve similar accuracy on FBK15 (or better)

checking on job 2414565

• Entity/Link Prediction (see #13)
• Triple Classification (see #12)