Hi,
Yay it's somebody else's bug. I changed KenLM to refuse (throw an
exception that cannot be disabled) for ARPA files that contain NaN
probability or backoff. It will also refuse infinite backoff weights.
In Moses, this is git 4bcd2c7.
Kenneth
On 03/11/2012 01:19 PM, Rico Sennrich wrote:
> Hi Ken, list,
>
> Thanks for having a look.
> The problem seems to be with the SRILM model:
>
> -1.534776 den haben . 0
> -3.705714 und den haben -0.6288081
> -2.604781 , und den haben nan
>
> I'd appreciate if you could add a warning during binarization, but maybe
> you're sick of dealing with broken models... :D
>
> @everybody else: it may be worth checking the backoff weights of your LM
> if you're getting occasional segfaults.
>
> best wishes,
> Rico
>
> _______________________________________________
> 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
