Hi, I trained my model with -phrase-word-alignment and created the phrase table with -alignment-info with a previous version.
If I run the decoder with: -use-alignment-info -print-alignment-info-in-n-best -n-best-list - 1 -alignment-output-file - -include-alignment-in-n-best I get the alignment information I want. I checked out the latest version today and installed the scripts. But if I compile a binary phrase-table with the -alignment-info switch (to preserve the word-to-word-alignment) and use it for tuning, I get the aforementioned error. The error does not occur, if I compile the binary phrase-table without the -alignment-info switch (the word-to-word-alignment gets lost) and use it for tuning. Regards, Andreas On Mon, Mar 14, 2011 at 6:13 PM, Philipp Koehn <[email protected]> wrote: > Hi, > > if you train your model with the switch "-phrase-word-alignment" to > train-model.perl > and run your decoder with "-alignment-output-file FILE", then you get the word > alignment information. Tuning should run without problems (or > additional switches). > > You need to check out the latest version from SVN, since there have been > very recent updates. > > -phi > > On Mon, Mar 14, 2011 at 11:26 AM, Andreas Kull <[email protected]> wrote: >> I'm referring to >> http://www.mail-archive.com/[email protected]/msg01917.html >> >> Perhaps I didn't understand it correctly. >> >> >> My error message: >> """ >> reading bin ttable >> size of OFF_T 8 >> ERROR:You are asking binary phrase table without word alignments but >> the file do not exist. Please check if you had generated the correct >> phrase table without word alignment >> (/home/user/mosesdecoder/corpora/europarl/50/training/model/phrase-table.binphr.srctree,/home/user/mosesdecoder/corpora/europarl/50/training/model/phrase-table.binphr.tgtdata) >> >> ERROR:bin ttable was read in a wrong way >> >> Segmentation fault >> Exit code: 139 >> """ >> >> The binary table is declared correctly and was created with the >> -alignment-info switch: >> >> In moses.ini: >> [ttable-file] >> 1 0 0 5 >> /home/user/mosesdecoder/corpora/europarl/50/training/model/phrase-table >> >> >> Regards, >> Andreas >> >> On Mon, Mar 14, 2011 at 11:04 AM, Philipp Koehn <[email protected]> wrote: >>> Hi, >>> >>> re-using tuning weights from a different run is perfectly fine. >>> >>> Having said that, I do not fully understand why it is not possible >>> to use the phrase table with word alignment information during >>> tuning. What seems to be the problem? >>> >>> -phi >>> >>> On Mon, Mar 14, 2011 at 9:05 AM, Andreas Kull <[email protected]> >>> wrote: >>>> Hi, >>>> >>>> as far as I know it is currently not possible to start the tuning >>>> process with binary phrase-tables that contain word alignment >>>> information (created via -alignment-info switch). >>>> >>>> Is it possible to use the binary phrase table without word-to-word >>>> alignment information for tuning and copy the generated tuning values >>>> in a moses.ini which uses a binary phrase-table with word-to-word >>>> alignment information? Is this a valid workaround or will I run into >>>> problems? >>>> >>>> >>>> Regards, >>>> Andreas >>>> _______________________________________________ >>>> Moses-support mailing list >>>> [email protected] >>>> http://mailman.mit.edu/mailman/listinfo/moses-support >>>> >>> >> _______________________________________________ >> Moses-support mailing list >> [email protected] >> http://mailman.mit.edu/mailman/listinfo/moses-support >> > _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
