i think you're asking why the unigram and bigram LM scores of the 1st two words are used to calculate future scores when the LM is a trigram. that's a good question & 1 i've revisited recently with the hierarchical moses. i'm not sure there's a good theoretical basis for it. however, the future score is also used to prune certain phrase pairs before decoding to speed up the process. Including the unigram and bigram score definately help in ensuring good translations aren't pruned. Hieu Hoang www.hoang.co.uk/hieu
_____ From: [email protected] [mailto:[email protected]] On Behalf Of Ergun Bicici Sent: 17 February 2009 13:10 To: Philipp Koehn Cc: [email protected] Subject: Re: [Moses-support] Future costs calculation in MOSES Hi Philipp, Thanks for the response. I was not asking why these scores are cached. My question is more about why calculate this way. Is this because of an admissible heuristic? Ergun Bicici Koc University On Wed, Feb 11, 2009 at 11:51 PM, Philipp Koehn <[email protected]> wrote: Hi, what is going here is a caching of phrase-internal n-gram model scores, so they do not have to be re-computed. Think about the output phrase "the very big and funny man" - if you use a trigram language model, then the computation of the language model scores for the words "big", "and", "funny", "man" are the same, no matter what the context. So, these are cached. -phi > LanguageModel::CalcScore is adding ngram score to retFull score: > fullScore += ngramScore; > > But then in TranslationOption::CalcScore, this is subtracted back: > m_futureScore = retFullScore - ngramScore > + > m_scoreBreakdown.InnerProduct(StaticData::Instance().GetAllWeights()) - > phraseSize * StaticData::Instance().GetWeightWordPenalty(); > > > - Is the n-gram order (3) fixed for LM cost calculations > used in future cost? It does not look so. > > > It would be helpful if someone could clarify the > future cost calculation further. > > Thanks, > Ergun > > > Ergun Bicici > Koc University > > > On Wed, Sep 24, 2008 at 5:46 PM, Philipp Koehn <[email protected]> wrote: >> >> Hi, >> >> the future cost estimates includes an estimate of the phrase translation >> cost >> and language model cost, but not reordering costs. And yes, this is >> implemented >> as described in the Pharaoh manual. >> >> -phi >> >> On Wed, Sep 24, 2008 at 8:58 AM, Yee Seng Chan <[email protected]> >> wrote: >> > Hi list members, >> > >> > >> > >> > Inside TranslationOption.cpp::CalcScore(), m_futureScore is effectively: >> > retFullScore - (PhraseSize*WordPenalty) >> > >> > (Kindly correct me if I'm wrong). >> > >> > >> > >> > What's the reasoning for using the above as futureScore? I know >> > retFullScore >> > is n-gram score. Btw, does the approach here follows "Section 3.5 Future >> > Cost Estimation" in the Pharaoh manual? >> > >> > >> > >> > Best regards, >> > >> > Yee Seng Chan >> > >> > >> > >> > _______________________________________________ >> > 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
