I am NOT talking about the [CORPUS:europarl] section but
the [LM:europarl] I think in this section you need the $pair-extension
same as [LM:nc] where it was fine
Anyway :yes I had an error
Europarl releases usually contain the parallel files (e.g.
europarl-v7.fr-en.fr) and monolingual files (e.g. europarl-v7.en).
If I am not mistaken, http://www.statmt.org/wmt12/training-parallel.tgz
do not, this could be the reason.
Also : when.gz is missing yes it stops and give an error message.
OK, this used to work.
Sorry I was not specific enough. the errors pops up when I use
./daemon.pl for the web translation. Could be just there ....
That's right, the interpolated LM uses some code from SRILM. You can
still use KenLM to create the individual language models, and use
KenLM during decoding,
OK. But related question :
If I do not interpolate, and if I keep the 2 (or more) LM in the
moses.ini file
does the decoder work similarly as if I had interpolated 2 LMs ?
If you do not interpolate using EMS then both LMs will be features in
the model - i.e. you get log-linear interpolation. See here for an
early comparison of linear and log-linear interpolation -
https://aclweb.org/anthology/W/W07/W07-0717.pdf - there has been other
work since then. Note that SRILM does not do linear interpolation
correctly,
Many thanks.
Just as a general question.
The baseline Tuto mentions this at the end :
This gives me a BLEU score of 23.5 (in comparison, the best result at
WMT11 was 30.5 <http://matrix.statmt.org/matrix/systems_list/1669>,
although it should be cautioned that this uses NIST BLEU, which does its
own tokenisation, so there will be 1-2 points difference in the score
anyway)
The Baseline tuto is done with NewsCommentary_V8. I did it and got 22/23
as a Bleu score.
My EMS with the config.basic file gives me around 26 (EuroparlV7+NCv10)
test set out of NC2011.
Is the reason for my "low" score versus 30s because I am using KenLM
only ? am I missing something else ?
cheers - Barry
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support
