search

artetxem / vecmap

A framework to learn cross-lingual word embedding mappings
GNU General Public License v3.0
645 stars 130 forks source link

Re-weighting difference between ACL2018 paper and code #13

Closed zhangxiangnick closed 6 years ago

zhangxiangnick commented 6 years ago

Really enjoyed your ACL2018 paper! Need some clarification for the re-weighting step.

In your ACL2018 paper, it says re-weighting is used as "a final step once self-learning has converged to a good solution."

In the code (line 314), it seems like re-weighting is used in every step of self-learning.

Does it make much difference how re-weighting is used?

artetxem commented 6 years ago

Re-weighting is only applied in the last iteration as described in the paper: when the variable 'end' is False (which happens in all iterations but the last one) we learn an orthogonal mapping (line 282).