Questions about the way to take use of the word embeddings. about muse HOT 1 CLOSED

NN (= Cosine Similarity) and CSLS (Cross-domain Similarity Local Scaling) are two measures of similarity between word vectors. If you have a source word in (say) Chinese, and you want to get its nearest neighbor in Latin, you may simply use the cosine similarity for simplicity (computing the cosine similarity with all the target words and retrieving the one with maximum similarity). For all the pairs that we have tried, CSLS provides better translation accuracy than cosine similarity, but also requires more computation. In your particular case (Chinese-Latin) it seems the results on word translation for NN or CSLS are similar, so you can safely use cosine similarity for simplicity. Note than Chinese-Latin is a difficult language pair (hence your 10% accuracy), be particularly careful of the quality of your word embeddings in Latin.

from muse.