Hi, if you specify multiple translations, you can associate a probability with each:
<x translation="house||home" prob="0.9||0.7"> haus </x> This number will be used for all translation model scores. Also, the language model will play a role, just as for normal phrase table entries. -phi On Mon, Dec 10, 2012 at 6:42 PM, Souhir Gahbiche <[email protected]> wrote: > Hi all, > About scores, another question : > In the case of multiple suggestions with xml and if the probabilities > are not given, for every suggestion, > - the decoder will not assign a probability but will looks in the > phrase table and language model if that phrase exists and give that > scores ? > OR > - the XML translation option is given a probability of 1 ? > > Bests > > 2012/4/3, Tom Hoar <[email protected]>: >> Thank you Philipp, this is a clear and convincing example of the power >> of the language model. >> >> Based on a previous discussion, I understood the xml markup using the >> -xml-input argument effectively supplements the phrase table with a >> temporary entry for that sentence translation. >> >> What affect does the reordering table have on the results using xml >> markup, if any, assuming the xml markup data is in the reordering table? >> >> >> Le 2 apr 2012 13:19, Philipp Koehn <[email protected]> a écrit : >>> note that the translation probability is not the only scoring >>> function influencing the translation decisions, there is also the >>> language model. >>> >>> To give an example: >>> >>> % echo 'die neuen <xml translation="gentlemen||councilmen" >>> prob="0.5||0.5">Rats Herren</xml> der Hauptstadt' || moses -f >>> moses.ini -xml-input exclusive >>> >>> the new gentlemen of the capital >>> >>> % echo 'die neuen <xml translation="gentlemen||councilmen" >>> prob="0.01||0.99">Rats Herren</xml> der Hauptstadt' || moses -f >>> moses.ini >>> -xml-input exclusive >>> >>> the new gentlemen of the capital >>> >>> % echo 'die neuen <xml translation="gentlemen||councilmen" >>> prob="0.00001||0.99999">Rats Herren</xml> der Hauptstadt' || moses -f >>> moses.ini -xml-input exclusive >>> >>> the new councilmen of the capital >>> >>> So, in this case (a recent WMT12-de-en system), it takes quite a bias >>> in >>> the translation model to produce the output with the higher >>> translation >>> probability. >>> >>> -phi >>> >>> On Mon, 2 Apr 2012 11:19:26 +0200, Souhir Gahbiche >>> <[email protected]> wrote: >>>> Hi, >>>> In the case when I use probabilities, the correct translation is the >>>> one which have the highest probability in the proposed options <n >>>> translation="Lutece||Paname||Pantruche||Ville-Lumière||Paris" >>>> prob=0.1||0.1||0.1||0.1||0.6>bArys</n> >>>> So, in this case, the correct suggestion is Paris. >>>> >>>> Regards >>>> S >>>> >>>> 2012/3/31, Souhir Gahbiche <[email protected]>: >>>>> Hi, >>>>> I mean by the "correct" suggestion, according to the reference. >>>>> The source word exist in the translation model and in the language >>>>> model. >>>>> I tried the "inclusive" mode, but the suggestion chosen doesn't >>>>> exist >>>>> neither in the translation model nor in the language model, and the >>>>> "correct" suggestion (the one which should be chosen) exists both >>>>> in tha >>>>> translation model and the language model. >>>>> >>>>> Bests >>>>> S >>>>> >>>>> Le 30 mars 2012 23:43, Philipp Koehn <[email protected]> a écrit >>>>> : >>>>> >>>>>> Hi, >>>>> >>>>>> could you clarify what you mean by "correct" suggestion. >>>>>> >>>>>> If the source word for which you provide XML options does not >>>>>> exist in the translation model, or if you use the "exclusive" >>>>>> setting, >>>>>> then the deciding factor here is the language model. If both >>>>>> options do not exist in the language model, then it's a coin flip. >>>>>> >>>>>> -phi >>>>>> >>>>>> On Fri, Mar 30, 2012 at 10:21 AM, Souhir Gahbiche >>>>>> <[email protected]>wrote: >>>>>> >>>>>> Hi all, >>>>>>> >>>>>>> I'm trying to decode using XML option, but I didn't get the >>>>>>> correct >>>>>>> suggestion when I propose many translations for the decoder. >>>>>>> In some cases of my test I got the correct suggestion, in another >>>>>>> cases >>>>>>> the same suggestion is not choosen. >>>>>>> >>>>>>> For example, for <n >>>>>>> >>>>>>> translation="Lutece||Paname||Pantruche||Ville-Lumière||Paris">bArys</n> >>>>>>> Ville-Lumière was chosen instead of Paris. It does not exist >>>>>>> neither in >>>>>>> the phrase-table nor in the language model. >>>>>>> >>>>>>> I tried inclusive/exclusive options, with and without >>>>>>> probabilities and >>>>>>> I >>>>>>> didn't obtain the correct suggestion. >>>>>>> The only case when I obtained the correct answer is when I give >>>>>>> to the >>>>>>> decoder only one suggestion. >>>>>>> >>>>>>> Regards >>>>>>> >>>>>>> _______________________________________________ >>>>>>> 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 >> > > _______________________________________________ > 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
