Thanks for your reply. Just to explain the general idea: I try to do some experiments with adaptive translation models using a simple cache. Some notion of confidence would help me deciding whether a certain translation option is reliable (and should be added to the cache) or not.
Actually, this model should better prefer content words and, therefore, it's probably not a good idea to use the hypothesis score for this decision. Maybe I should just rely on the translation scores from the phrase table? Or I could at least leave out (subtract) the LM score? Bit in some sense it would be nice to include contextual knowledge instead of simply using the individual phrase translation scores. Would there be any other way to bias towards content words? Some kind of length normalization maybe? Any thoughts? Jörg > > yes, GetScore() does not include the future score, just the weighted > partial score computed so far. The difference to > hypo->GetPrevHypo()->GetScore() > is the transition cost from the prior best hypothesis. This number should > be negative, excecpt in rare cases (for instance adding a common word > that the language model does not mind, but the brevity penalty rewards). > > I am not sure what you are planning to do with "confidence", but bear in mind > that longer phrases and phrases with less common (content) words > will be more costly. > > -phi > > 2010/2/3 Jörg Tiedemann <jorg.tiedem...@lingfil.uu.se>: >> I have a question about scores for individual translation hypotheses. >> I would like to get some kind of confidence measure for a particular >> phrase translation used in a given translation hypothesis. >> >> Does it make sense to compute the difference between the current score >> (using hypo->GetScore()) and the score at the previous step >> (hypo->GetPrevHypo()->GetScore())? This ignores any future score >> estimations if I understand correctly but that should be OK, shouldn't >> it? (I don't want to include any other costs here because I'm only >> interested in the confidence of the current hypo within the translation >> candidate.) By the way, this difference can be a positive value, right? >> It only depends on the weights and feature values, doesn't it? But would >> it make sense to have a fixed threshold for defining some level of >> confidence for this candidate? Probably not ....?! >> >> Thanks for any comments, >> >> Jörg >> >> *******/\/\/\/\/\/\/\/\/\/\/\****************************************** >> Jörg Tiedemann jorg.tiedem...@lingfil.uu.se >> http://stp.lingfil.uu.se/~joerg/ >> Dep. of Linguistics and Philology >> Uppsala University tel: +46 (0)18 - 471 1412 >> Box 635, SE-751 26 Uppsala/SWEDEN fax: +46 (0)18 - 471 1094 >> *********************************/\/\/\/\/\/\/\/\/\/\/\**************** >> _______________________________________________ >> Moses-support mailing list >> Moses-support@mit.edu >> http://mailman.mit.edu/mailman/listinfo/moses-support >> -- *******/\/\/\/\/\/\/\/\/\/\/\****************************************** Jörg Tiedemann jorg.tiedem...@lingfil.uu.se Visiting Professor http://stp.lingfil.uu.se/~joerg/ Dep. of Linguistics and Philology Uppsala University tel: +46 (0)18 - 471 1412 Box 635, SE-751 26 Uppsala/SWEDEN fax: +46 (0)18 - 471 1094 *********************************/\/\/\/\/\/\/\/\/\/\/\**************** _______________________________________________ Moses-support mailing list Moses-support@mit.edu http://mailman.mit.edu/mailman/listinfo/moses-support