Hi Lefty
The tokeniser was updated at some point so that it escaped certain
characters such as single quote ('). Could this be the problem? If you
tokenise your input differently from the training data you will get bad
results,
cheers - Barry
On 29/07/13 19:26, Hieu Hoang wrote:
that's very worrying. This is a major reason we're running end-to-end
training as part of the releases, to make sure the whole pipeline
hasn't degraded.
a few questions:
1. is the test or tuning data in any way unstable? eg. it's really
small. Can you run the entire training again and see what the bleu it
get again.
2. with the same weights as the 2011 version, does the decoder get
the same model and bleu score?
3. Can you make the 2011 working data available for download and
i'll take a look
On 29 July 2013 17:45, Eleftherios Avramidis
<[email protected] <mailto:[email protected]>>
wrote:
Hi,
I have build some models some time ago, using moses and training code
which had been checked out in 2011. Now I need to improve these
models,
so I decided to rerun the old configuration with the freshly
checked-out
Moses. I realize that my BLEU scores, on the exact same configuration,
are down by 0.5 to 1.2 BLEU score.
Both were trained/tested with the same version of experiment.perl.
I guess I will have to either dig to find what exactly made the
difference, or just use the old version if I need improvements. But
would you think an obvious reason why this happened?
cheers
Lefteris
--
MSc. Inf. Eleftherios Avramidis
DFKI GmbH, Alt-Moabit 91c, 10559 Berlin
Tel. +49-30 238 95-1806 <tel:%2B49-30%20238%2095-1806>
Fax. +49-30 238 95-1810 <tel:%2B49-30%20238%2095-1810>
-------------------------------------------------------------------------------------------
Deutsches Forschungszentrum fuer Kuenstliche Intelligenz GmbH
Firmensitz: Trippstadter Strasse 122, D-67663 Kaiserslautern
Geschaeftsfuehrung:
Prof. Dr. Dr. h.c. mult. Wolfgang Wahlster (Vorsitzender)
Dr. Walter Olthoff
Vorsitzender des Aufsichtsrats:
Prof. Dr. h.c. Hans A. Aukes
Amtsgericht Kaiserslautern, HRB 2313
-------------------------------------------------------------------------------------------
_______________________________________________
Moses-support mailing list
[email protected] <mailto:[email protected]>
http://mailman.mit.edu/mailman/listinfo/moses-support
--
Hieu Hoang
Research Associate
University of Edinburgh
http://www.hoang.co.uk/hieu
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support