Lane - if it's not reachable, it should give you a blank line and 'NO 
BEST TRANSLATION' on stderr. Additonally, there won't be anything in the 
nbest list. I don't think i kept your multi-reference implementation, my 
have to redo that ;) git pull, it might have, err, crashed on the 1st 
iteration.

Ondrej - it's applied during the search, when each hypothesis is added 
to the stack. We can think about pruning the translation options before 
the search

If the hypothesis doesn't create a substring of the reference, the 
feature function will give it a score of -infinity and the stack will 
immediately discard it.

I think it's a flexible feature, eg. you can replace the exact match 
with word edit distance etc.

Christian - allowing some unknown word is a good idea. I've added a 
'max-unknowns' argument to the feature function, eg
    [feature]
    ...
    ConstrainedDecoding path=ref.txt max-unknowns=2

this is 0 by default, -1 makes it allow any number of unknowns words.

On 18/09/2013 11:10, Ondrej Bojar wrote:
> Hi,
>
> great! Thanks!
>
> When is the constraint applied? After translation options have been created, 
> or already during their creation? I guess the former, since the latter is 
> somewhat difficult to apply.
it works during search,
> My main concern is the relationship between constraint decoding and factored 
> setups. I guess the current version definitely works with alternative 
> decoding paths of single-factored translation, and it also probably works 
> with any factored setup *modulo* pruning of translation options. (If the 
> needed translation option is available but happens to get pruned, the 
> sentence will fail.)
>
> Please correct me if I am wrong.
>
> Cheers, O.
>
> ----- Original Message -----
>> From: "Hieu Hoang"<[email protected]>
>> To: "Maria Nadejde"<[email protected]>, "Lane 
>> Schwartz"<[email protected]>, "Ondrej Bojar"
>> <[email protected]>, "moses-support"<[email protected]>
>> Sent: Wednesday, 18 September, 2013 10:50:28 AM
>> Subject: constrained decoding
>>
>> For anyone who's interested, the constrained decoding has been
>> reimplemented as a feature function and works for both phrase-based and
>> syntax models.
>>
>> To constrain the output of the decoder to just the reference sentences, add
>> this as a feature:
>>     [feature]
>>      ....
>>      ConstrainedDecoding path=ref.txt
>>
>>
>> --
>> Hieu Hoang
>> Research Associate
>> University of Edinburgh
>> http://www.hoang.co.uk/hieu
>>

_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to