word-embeddings-benchmarks icon indicating copy to clipboard operation
word-embeddings-benchmarks copied to clipboard

Published 20 hours ago • verified publisher icon kudkudak

SG Google News embeddings (word2vec) are reaching 0.726 on Google word analogy corpus

Open hanselowski opened this issue 6 years ago • 2 comments

Thanks a lot for providing this code, this helps for a fair comparison of frameworks. According to my experiments with your tool, SG Google News embeddings (word2vec) are reaching a performance of 0.726 on the google word analogy corpus and not 0.402. The small number seems unrealistic. Thanks!

hanselowski avatar Nov 22 '18 13:11 hanselowski

Have you tried normalizing the word vectors before passing them to the analogy solver? This could help

kudkudak avatar Nov 25 '18 06:11 kudkudak

No I mean the table on your GitHub repository seems wrong to me. I think SG Google News embeddings (word2vec) are reaching a performance of 0.726 on the google word analogy corpus and not 0.402 as in the table.

hanselowski avatar Nov 25 '18 11:11 hanselowski