nmrksic / counter-fitting Goto Github PK

Hello,

Is there code available or an explanation somewhere on the data processing? Specifically, I'm interested in how you generated ppdb-synonyms.txt and wordnet-antonyms.txt.

Thanks!

the partial derivative of AR，SA，VSP are computed as follows：
if normalised_vectors:
gradient = u * dot(u,v) - v

is that right？the only parameter is v'，right？ both v'i and v'j should be computed partial derivatives，right？

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.

I compared the vector values in glove.txt with corresponding vectors from http://nlp.stanford.edu/projects/glove/. However, for the same words, I couldn't find matching vectors.

Could you let me know where did you obtain the vector values of glove.txt?

Thanks!

Joo-Kyung Kim

Hi,

Thanks for your dedication of this work!

I think there are some typos in the generated embedding file. For example, 'recieve' and 'recieved', while 'receive' and 'received' also appear. Besides, I guess there are some words such as 'worden' and 'viens' which look like vocabularies taken from other languages other than English. It could be a problem when I want to generate synonyms to replace words in the original sentences.

Do you have any idea to avoid these words?

First, thanks so much for posting your code!

Counter-fitting glove vectors with a much larger vocab (~2 million words) the number of required dot products for computing VSP pairs obviously explodes (2m^2 vs 50k^2). The included pruned vocab of 50k only takes ~1min on a 4 core machine. What type of time frame did your full glove vocab take? I feel like I must be missing something.

Thanks again Nikola!