Hi,

if not all of your text is wrapped in SGML then there is a problem.
Since I do not know about the nature of your sets, it's hard to explain
why that happens, but you should know why the sentences in the
SGML file do not correspond to the system output.

Tuning on larger (and hence different) tuning sets may decrease
BLEU scores on a particular test set, there are really no guarantees.

-phi

On Sun, Aug 10, 2008 at 1:59 AM, Vineet Kashyap
<[EMAIL PROTECTED]> wrote:
> Hi all
>
> I think that this problem is fixed as i am getting
> BLEU scores but was wondering is this the right way of wrapping
> these files in xml. What is the example-set?
>
> When i use ref.sgm src.sgm and tst.sgm files as refernce to produce
> actual files in sgm i get only a portion of the testing and output
> wrapped. Will this affect the BLEU scores?
>
> The scores i am getting for a very small corpus are as follows:
>
> NIST score = 3.9406  BLEU score = 0.1323 for system "ref1"
> NIST score = 4.9375  BLEU score = 0.1776 for system "ref2"
> NIST score = 5.2981  BLEU score = 0.4421 for system "ref3"
> NIST score = 5.4734  BLEU score = 0.2691 for system "ref4"
>
> Also strange thing is that after optimization on a tuning set of 5000
> sentences the BLEU scores go down
>
> Is there something i am missing?
>
> Thanks
>
> Vineet
>
>
>
>
>
>
> _______________________________________________
> 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

Reply via email to