tech,8-1-P05-1069,ak In this paper , we present a novel <term> training method </term> for a <term> localized phrase-based
model,12-1-P05-1069,ak novel <term> training method </term> for a <term> localized phrase-based prediction model </term> for <term> statistical machine translation
tech,17-1-P05-1069,ak phrase-based prediction model </term> for <term> statistical machine translation ( SMT ) </term> . The <term> model </term> predicts <term>
model,1-2-P05-1069,ak machine translation ( SMT ) </term> . The <term> model </term> predicts <term> blocks with orientation
other,3-2-P05-1069,ak </term> . The <term> model </term> predicts <term> blocks with orientation </term> to handle <term> local phrase re-ordering
tech,8-2-P05-1069,ak blocks with orientation </term> to handle <term> local phrase re-ordering </term> . We use a <term> maximum likelihood
measure(ment),3-3-P05-1069,ak phrase re-ordering </term> . We use a <term> maximum likelihood criterion </term> to train a <term> log-linear block
model,9-3-P05-1069,ak likelihood criterion </term> to train a <term> log-linear block bigram model </term> which uses <term> real-valued features
other,15-3-P05-1069,ak block bigram model </term> which uses <term> real-valued features </term> ( e.g. a <term> language model score
measure(ment),20-3-P05-1069,ak real-valued features </term> ( e.g. a <term> language model score </term> ) as well as <term> binary features
other,27-3-P05-1069,ak language model score </term> ) as well as <term> binary features </term> based on the <term> block identities
other,32-3-P05-1069,ak binary features </term> based on the <term> block identities </term> themselves , e.g. <term> block bigram
other,37-3-P05-1069,ak identities </term> themselves , e.g. <term> block bigram features </term> . Our <term> training algorithm </term>
tech,1-4-P05-1069,ak <term> block bigram features </term> . Our <term> training algorithm </term> can easily handle millions of <term>
other,8-4-P05-1069,ak </term> can easily handle millions of <term> features </term> . The best system obtains a 18.6
other,10-5-P05-1069,ak obtains a 18.6 % improvement over the <term> baseline </term> on a standard <term> Arabic-English
other,14-5-P05-1069,ak <term> baseline </term> on a standard <term> Arabic-English translation task </term> . Despite much recent progress on
skrij podrobnosti