I was trying to look at the task of link prediciton on the WNR11 dataset. I looked at the TuckER model and found that the TuckER model can do very well on the facts involving symmetric relation types. The model score around 0.9 on Hits@10,Hits@3, Hits@1, and MRR. However restricting the dataset to be only the antisymmetric relation data types then the model perform much worse. It score less then 0.2 on all 4 scoring metric(Hits@10,Hits@3, Hits@1 and MRR).
I'm looking at the code for TuckER and I'm wondering what are possible techniques to improve the model?