Skip to Main Content
Table 4: 
bleu search error, and average number of calls to score for output obtained with length normalization scoring function on the IWSLT’14 De-En and MTTT Fr-En test sets. Increase in bleu is over baseline with no length normalization. Search error and performance increases are with respect to standard beam search decoding using the same scoring function.
IWSLT’14 De-En
kβb# callssearch errorbleu
Heuristic 0.8 |x115(0%) 40.6% 33.9+0.3 
10 1.2 |x229(0%) 54.7% 33.8+0.5 
 
Stopping Criterion 0.5 nmax 73(58%) − 33.7+0.1 
10 0.5 nmax 130(76%) − 33.7+0.4 
 
MTTT Fr-En 
Heuristic 0.8 .7|x100(8%) 16.2% 33.5+0.2 
10 1.0 .7|x196(9%) 25.2% 33.6+0.6 
 
Stopping Criterion 1.0 nmax 65(66%) − 34.1+0.8 
10 1.2 nmax 88(143%) − 34.1+1.1 
IWSLT’14 De-En
kβb# callssearch errorbleu
Heuristic 0.8 |x115(0%) 40.6% 33.9+0.3 
10 1.2 |x229(0%) 54.7% 33.8+0.5 
 
Stopping Criterion 0.5 nmax 73(58%) − 33.7+0.1 
10 0.5 nmax 130(76%) − 33.7+0.4 
 
MTTT Fr-En 
Heuristic 0.8 .7|x100(8%) 16.2% 33.5+0.2 
10 1.0 .7|x196(9%) 25.2% 33.6+0.6 
 
Stopping Criterion 1.0 nmax 65(66%) − 34.1+0.8 
10 1.2 nmax 88(143%) − 34.1+1.1 
Close Modal

or Create an Account

Close Modal
Close Modal