On 15 March 2014 19:58, Peter Kleiweg <[email protected]> wrote:

>
> Hi all,
>
> I am trying to do factored translation.
>
> I have an input language (en) with these factors:
>
>   0. word
>   1. lemma
>   2. pos
>   3. rel
>
> I have a target language (nl) with these factors:
>
>   0. word
>   1. lemma
>   2. pos
>   3. postag
>   4. rel
>
> TRAINING
>
> I want to use word(0), lemma(1) and pos(2) from source language (en),
> and word(0), lemma(1), postag(3) from target language (nl).
>
> I run this command:
>
> nohup /net/aps/64/opt/moses/mosesdecoder/scripts/training/train-model.perl
> \
>   -cores 12 \
>   -root-dir train \
>   -corpus
> /net/aistaff/kleiweg/moses/europarl/factored/corpus/train-small.clean \
>   -f en -e nl -alignment grow-diag-final-and -reordering
> msd-bidirectional-fe \
>   --lm 0:3:/net/aistaff/kleiweg/moses/europarl/factored/lm/blm.nl-word:8 \
>   --lm 3:3:/net/aistaff/kleiweg/moses/europarl/factored/lm/blm.nl-postag:8
> \
>   --translation-factors 1-1+2-3+0-0,3 \
>   --generation-factors 1-3+1,3-0 \
>   --decoding-steps t0,g0,t1,g1:t2 \
>   -external-bin-dir /net/aps/64/opt/moses/mgizapp/bin \
>   -mgiza \
>   -mgiza-cpus=12 \
>    > training.out \
>    2> training.err &
>
> Did I specify the right translation-factors, generation-factors
> and decoding-steps?
>
That looks ok, there's nothing that contradict each other.

whether it's 'right' depends on your point of view. Personally, I try to
make the factored models as easy as possible. The complicated models with
multiple translation models & generation models make too much independence
assumptions.


> Did I specify the right language models?
>
it should run

>
> Should I specify --alignment-factors, like this:
>
>   --alignment-factors 1,2-1,3
>
up to you

>
> Training ends with these lines on stderr:
>
> (9) create moses.ini @ Sat Mar 15 17:19:20 CET 2014
> Use of uninitialized value $___REORDERING_FACTORS in split at
> /net/aps/64/opt/moses/mosesdecoder/scripts/training/train-model.perl line
> 1997.
>
> Should I specify --reordering-factors ? What does this do? What
> do I specify? All I can find in the manual is this:
>
> "Reordering tables can be trained with --reordering-factors.
> Syntax is the same as for translation factors."
>
> That isn't very helpful.
>
> it  specifies the factors to use when training the lexicalised reordering
model
   http://www.statmt.org/moses/?n=FactoredTraining.BuildReorderingModel

The model that was created seems to be working (even though step
> 9 of training gave an error), but it's slow, 10 minutes or more
> for a sentence, where unfactored translation only took a few seconds.
>
because you have a complicated factored model. It does a cross-product of
all the translation model & generation models

>
> TUNING
>
> What type of data should I give mert-moses.pl to work with? I
> noticed the source language should only have factors 0, 1 and 2,
> or I get an error. What factors should there be in the target
> language?
>
>
>
> --
> Peter Kleiweg
> http://pkleiweg.home.xs4all.nl/
> _______________________________________________
> Moses-support mailing list
> [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

Reply via email to