Hello Prashant,

The total drop in BLEU should be a maximum a couple BLEU points. If it's
more than this there
will be an error in the training setup or the data isn't correctly
formatted.

Fixing this discrepancy in the BLEU score is an area that we should look
into shortly.

Abby


On Thu, Apr 11, 2013 at 5:52 PM, Prashant Mathur <[email protected]> wrote:

> Hello All,
>
> I am running some experiments using dynamic suffix array in Moses.
> for that, I use this
> 8 0 0 3 source-corpus target-corpus alignment
> instead of standardized
> 0 0 0 5 phrase-table
>
> Since, using suffix array generates only 3 scores.. I am wondering how
> much drop in performance in respect to standard system did others see in
> their experiments?
>
> On one of the dataset that I am using the I can see a drop of 10 BLEU
> points, whereas on another dataset the drop is only of 2-3 points.
>
> I believe the drop in BLEU points should be consistent across various
> corpus. Can anyone comment a bit more on this?
>
> Thanks,
> Prashant
>
>
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to