it is reasonable？

[bringtree]

counter-fitting/counterfitting.py

Line 375 in 886b7b8

update_term = gradient_updates[word] / (update_count[word])

I think that the count should be split to 3 types[AR,SA,VSP] to compute.
Then the k1,k2,k3 weight can be work effectivity.

[lshowway]

@bringtree I agree with you, the update_term may be not reasonably computed。
do you know why the partial derivative is as follows？
gradient = u * dot(u,v) - v

[bringtree]

@lshowway

The vector had been done normalization.
It is a unit vector。

[lshowway]

@lshowway

The vector had been done normalization.
It is a unit vector。

Thank you. But, why the partial derivative is as follows？
gradient = u * dot(u,v) - v