Thanks Andreas, Yes, that helps. I'd tried it before I sent the mail, but only without n-best-list 1.
$ echo "Unix is a registered trademark licensed through X / Open Company , Ltd ." | TMP=/tmp /export/home/moses/export/home/mosesmt/moses/moses-cmd/src/moses -f /export/home/moses/engines/en2de/toy_alignment/ini/regression_tiny_alignment/moses.ini -include-alignment-in-n-best true -use-alignment-info -print-alignment-info-in-n-best -n-best-list - 1 -alignment-output-file - -include-alignment-in-n-best ...... Translating: Unix is a registered trademark licensed through X / Open Company , Ltd . Collecting options took 0.000 seconds Search took 0.130 seconds 0 ||| Unix ist eine eingetragene Marke qualifizierten ber X / Open Company , Ltd . ||| d: 0 -3.20347 0 0 -1.88542 0 0 lm: -60.1023 w: -14 tm: -4.12539 -5.46622 -1.05002 -4.65647 6.99927 ||| -419.238 ||| 0=0 1-4=1-4 5=5 6=6 7=7 8=8 9=9 10=10 11=11 12=12 13=13 ||| // end_of_output I am not sure if "1-4=1-4" is a right word-to-word alignment information. I expected a mapping for individual English words, such as "1=1" "2=2". Do you think this is an expected output? It doesn't look so if I compare the output with I am using non-binary phrase table just in case. And the phrase table looks like this. ) cable is ||| - Kabel ) ist ||| 1 0.25041 1 0.0454642 2.718 ||| 1-0 1-1 0-2 2-3 ||| 1 1 ) cable ||| - Kabel ) ||| 1 0.32136 1 0.118965 2.718 ||| 1-0 1-1 0-2 ||| 1 1 ) in the ||| ) im Feld ||| 1 0.434847 1 0.00463835 2.718 ||| 0-0 1-1 2-2 ||| 1 1 ) in ||| ) im ||| 1 0.640827 1 0.316384 2.718 ||| 0-0 1-1 ||| 1 1 FYI, the latest checkout from SVN was Feb. Thanks Jung On 16 March 2011 09:33, <[email protected]> wrote: Re: [Moses-support] Word alignment > > Andreas Kull > Wed, 16 Mar 2011 02:36:37 -0700 > > Hi, > > I hope that I understood your question. I don't know about the score > tool but regarding the decoder and the training options: > > 1. Training: You must use train-model.perl with the > -phrase-word-alignment switch > > 2. Decoder output: To get the word and phrase alignment, you will need > to specify the following switches for /moses-cmd/src/moses: > > -use-alignment-info > -print-alignment-info-in-n-best > -n-best-list - 1 > -alignment-output-file - > -include-alignment-in-n-best > > With this you will get the phrase and word alignment to stdout. > > > I hope this helped, > Andreas > > Date: Wed, 16 Mar 2011 09:25:59 +0000 > From: "Jungwoo(Nick) Ryoo" <[email protected]> > Subject: [Moses-support] Word alignment > To: [email protected] > Message-ID: > <[email protected]> > Content-Type: text/plain; charset="iso-8859-1" > > Hi, > > I hope I am repeating questions someone already asked. I am trying to get > the word alignment information using the toy model before rebuilding the > real model for that purpose. > > 1) It seems I need to give "phrase-word-alignment" option in > train-model.perl to put the alignment information in phrase table. But it > is > not documented anywhere. > > 2) I suppose printing word-alignment from moses decoder is still disabled. > I > couldn't print out the information. > > I tried to follow the advise from Philip. > http://www.mail-archive.com/[email protected]/msg01740.html > > But I couldn't use "--WordAlignment FilePath" option The "score" tool > didn't > like it with "unknown option" error. > > $ > > /export/home/moses/moses-scripts/current/training/phrase-extract/score/export/home/moses/engines/shared/en2de/models/toy_alignment/truecased_model/model/extract.sorted > > /export/home/moses/engines/shared/en2de/models/toy_alignment/truecased_model/model/lex.f2e > > /export/home/moses/engines/shared/en2de/models/toy_alignment/truecased_model/model/phrase-table.half.f2e > --WordAlignment > > /export/home/moses/engines/shared/en2de/models/toy_alignment/truecased_model/model/ > wordalignment.info > > Score v2.0 written by Philipp Koehn > scoring methods for extracted rules > outputing word alignment > ERROR: unknown option > /wptg/engines/shared/en2de/models/toy_alignment/truecased_model/model/ > wordalignment.info > Exit code: 1 > > But the usage clearly says that it accepts such pattern of parameters > though. > > $ > > /export/home/canaan/wptgmt/moses-scripts/current/training/phrase-extract/score > Score v2.0 written by Philipp Koehn > scoring methods for extracted rules > syntax: score extract lex phrase-table [--Inverse] [--Hierarchical] > [--OnlyDirect] [--LogProb] [--NegLogProb] [--NoLex] [--GoodTuring] > [--WordAlignment file] > > I wonder if there is any other way to get the word alignment information. > > Jung@Oracle. >
_______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
