tag | 1b1919b648e60e3c4e31e16adb9e82f411b7e238 | |
---|---|---|
tagger | Matt Post <post@cs.jhu.edu> | Wed Dec 17 15:56:08 2014 -0500 |
object | 23e7e7b53a78d26b08d57be3d44f33802c462bc5 |
6.0 release candidate 1 (not ready)
commit | 23e7e7b53a78d26b08d57be3d44f33802c462bc5 | [log] [tgz] |
---|---|---|
author | Matt Post <post@cs.jhu.edu> | Wed Dec 17 15:35:36 2014 -0500 |
committer | Matt Post <post@cs.jhu.edu> | Wed Dec 17 15:35:36 2014 -0500 |
tree | aa1b2db48732f1369c227b028743255425900bfb | |
parent | 54ebd2b5b99ceaf1861629d72996bef5ca9bb953 [diff] |
The OOV penalty is now -100 inside the decoder instead of 1 This means that the decoder itself always strongly disprefers OOVs; tuning can still downweight it, but this is more natural than relying on a hugely negative weight for OOVs