Hi Lane,
        Such constraint decoding function is interesting. Before several
days, I have tried this "-constraint" setting of moses. It can do force
decoding mostly with one bug.
        For example, if target sentence is "ABCDE". I have got one forced
translation by "ABCD". I guess it is caused by no length limit of translated
sentence by the last hypothesis.
        Anyway, such functionality is very useful. Thank you Lane.

Best regards,
Lang Jun

2010/12/2 Lane Schwartz <dowob...@gmail.com>

> Hi all,
>>
>> I'm trying to figure out if Moses, in its current state, allows for
>> forced decoding?
>>
>> I've understood that Lane Schwartz implemented it for his multi-source
>> translation (Schwartz, 2008)
>> and it is used by Auli et al. (2009). Is this implementation merged back
>> to the trunk?
>> If not in which branch could it be found?
>>
>> Any guidance would be highly appreciated.
>>
>> Cheers,
>> Nadi
>>
> Nadi,
>
> The code that I used for my multi-source translation paper (Schwartz, 2008)
> is in trunk (since r1857, as per the footnote in the paper). To use, specify
> a file of existing translations using the -constraint flag. This file needs
> to have one sentence per line, and should be the same number of lines as
> your input.
>
> The constraint decoding implemented for -constraint is pretty restrictive.
> If it can't find an path through the translation lattice that exactly
> produces the constraining sentence, it will fail to produce any translation
> for that sentence. In other words, if you specify "the man walks there" as
> your constraint, and the best the decoder can find is "the man walk there",
> translation will fail for that sentence.
>
> I did some work at the Prague MT Marathon to try and fix this issue, but
> that code is not in trunk, and I haven't touched it in a long time. Most of
> that code should be in the lane_mtm3 branch, but some also may be in the
> mt3_constraint branch (Hieu helped by implementing a WER component),
>
> Let me know if you plan to do any coding to extend constraint decoding.
>
> Cheers,
> Lane
>
>
> _______________________________________________
> Moses-support mailing list
> Moses-support@mit.edu
> http://mailman.mit.edu/mailman/listinfo/moses-support
>
>
_______________________________________________
Moses-support mailing list
Moses-support@mit.edu
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to