I have managed to replicate the results on the paper for English and German FastText. For reference, I am interested in the cross-lingual word similarity task. Results I got are (reporting spearman correlation):
Original FastText: 9%
Mapped FastText: 71%
However, I tried the same code on English and German GloVe embeddings and did not get much improvement. Results:
Original GloVe: 1%
Mapped GloVe: 3%
I have managed to replicate the results on the paper for English and German FastText. For reference, I am interested in the cross-lingual word similarity task. Results I got are (reporting spearman correlation): Original FastText: 9% Mapped FastText: 71%
However, I tried the same code on English and German GloVe embeddings and did not get much improvement. Results: Original GloVe: 1% Mapped GloVe: 3%
Any idea why this might be the case?