Hi,

there is also that a larger language model leads to state splitting
during decoding, so you may have to try larger beam sizes to
see gains.

-phi

On Wed, Mar 17, 2010 at 8:14 PM, Hieu Hoang <[email protected]> wrote:

> hi somayeh
>
> if you change the LM order, you should change the ini file so the decoder
> knows the order of the LM. eg an LM of order 55:
>    [lmodel-file]
>    0 0 55 /path/lm.srilm.gz
> it should go up. However, mert is random so its not guaranteed. Also, a
> higher order LM may make the decoder overfit the data it was trained on, so
> performance when translating unseen text may decrease. Espeically for
> out-of-domain text
>
> On 17 March 2010 18:55, Somayeh Bakhshaei <[email protected]> wrote:
>
>> Hello everybody
>>
>> How it is may that i increase the LM order, but the Bleu score of the
>> output decreases?!
>> If i use a higher LM order should i mention it explicitly in the train
>> part, phrase extraction , evaluation , ... ?
>>
>>
>> ------------------
>> Best Regards,
>> S.Bakhshsaei
>>
>>
>>
>> _______________________________________________
>> 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
>
>
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to