Hi,

the error stems from the low weight given to one of the
language models (4.78862e-05) which is not properly
parsed by the script scripts/ems/support/interpolate-lm.perl
which pattern matches weights with /best lambda \(([\d\. ]+)\)/

The simple solution here is to remove the first language model
since it does not add any value.

-phi

On Thu, Jul 28, 2011 at 10:12 AM, Sriram venkatapathy
<[email protected]> wrote:
> Hello,
>
> I am using multiple language models for my experiments, and am using
> Interpolated-LM to optimize the perplexity on the tuning set.
>
> It had worked well when I used a particular tuning set (set A) . But
> when I used a different one (set B), it crashed. Here is the log of
> the error:
>
> -----
> Executing: /home/svenkata/tools/srilm/bin/i686//compute-best-mix
> /tmp/dhDarpsIUB/iplm.32687.0 /tmp/dhDarpsIUB/iplm.32687.1
> /tmp/dhDarpsIUB/iplm.32687.2 /tmp/dhDarpsIUB/iplm.32687.3
> ERROR: computing lambdas failed: iteration 1, lambda = (0.5 0.166667
> 0.166667 0.166667), ppl = 28.1782
> iteration 2, lambda = (0.149855 0.0787999 0.349265 0.42208), ppl = 15.8558
> iteration 3, lambda = (0.0425529 0.0338758 0.394179 0.529393), ppl = 14.0577
> iteration 4, lambda = (0.0133561 0.016913 0.398085 0.571646), ppl = 13.6439
> iteration 5, lambda = (0.00464723 0.0099183 0.394268 0.591166), ppl = 13.5249
> iteration 6, lambda = (0.00174248 0.00659117 0.390115 0.601551), ppl = 13.4844
> iteration 7, lambda = (0.000684944 0.00477971 0.386928 0.607607), ppl = 
> 13.4687
> iteration 8, lambda = (0.000277308 0.00367993 0.384704 0.611339), ppl = 
> 13.4619
> iteration 9, lambda = (0.000114464 0.00295462 0.383217 0.613713), ppl = 
> 13.4586
> iteration 10, lambda = (4.78862e-05 0.00244556 0.382255 0.615251), ppl = 
> 13.4569
> iteration 11, lambda = (2.02318e-05 0.00207096 0.381653 0.616256), ppl = 
> 13.456
> 1422 non-oov words, best lambda (8.61264e-06 0.00178499 0.381292 0.616914)
> -----
>
> Any suggestions about what the issue might be ? The problem may not be
> with the tuning 'set B' itself, as I had used it with a different
> training conditions and there was no problem.
>
> Thanks !
> Sriram
> _______________________________________________
> 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