D09-1008 trained on the English side of the parallel training data . For that purpose , we
D08-1078 2000 sentences from each of the parallel training corpora . From this subset we
D09-1040 come to mind is to use larger parallel training corpora . However , current state-of-the-art
C96-1033 independently of each other , allowing for parallel training on several CPU 's . In run mode
D09-1074 sentence - aligned , word-aligned parallel training data , one could extract various
D08-1066 word-level alignments for the parallel training corpus using GIZA + + . We use
D09-1074 is easy to see that most of the parallel training data are either newswire or from
D08-1066 function of the word-aligned , parallel training corpus . Earlier efforts on devising
D09-1040 are limited by the quantity of parallel training texts . Augmenting the training
D09-1123 to parse the target side of the parallel training data . Each sentence is associated
D09-1074 probabilities . Typically , a parallel training corpus is comprised of collections
D10-1041 covered by the phrase table and the parallel training data . Section 3 describes our
D08-1090 4.2 used only 5 million words of parallel training , 230 million words of parallel
D08-1078 Each language pair has a separate parallel training corpus , but the target vocabulary
D09-1074 weight for each sentence in a parallel training corpus so as to optimize MT performance
D08-1090 condition involved a small amount of parallel training , such as one might find when
D09-1040 source and target sides of the parallel training sets . When the baseline system
D09-1074 identify , for each sentence in the parallel training data , a set of features that
D08-1090 versus 0.68 % BLEU . 4.3 Full Parallel Training Results While the simulation
D09-1040 rate concentrates on increasing parallel training set size without using more dedicated
hide detail