The signal matrix for word2vec

[XuhuiZhou]

In your paper, you seem to use the original PMI matrix as the signal matrix of word2vec. However, in Omer's paper, their conclusion points to the signal matrix PMI-log(k), where the k is the num of negative sampling.

Could you tell me is there a reason to choose PMI instead of PMI-log(k)?

[ziyin-dl]

As we said in the paper, we used PPMIfor word2vec without neg sampling, and SPPMI for word2vec with negative sampling. This repo is not for the one in the paper. You probably want to refer to
https://github.com/ziyin-dl/word-embedding-dimensionality-selection