Some more hints: there have been two papers in the last MT Summit about
word alignment features that may help SMT systems ("Reassessment of the
Role of Phrase Extraction in PBSMT" and "Tracking Relevant Alignment
Characteristics for Machine Translation"). In any case, I agree with Chris
that the impact on translation of word alignment differences is usually small.
On
relatively small corpora, directly tuning the word alignment system
according to MT criteria is another modestly successful example (see
Lambert Naacl 07 paper or the MT Summit paper).
Patrik
> There has been lots of work trying to improve word alignment so as to
> improve the quality of phrase translation models that depend on it.
> But, in the end, many of the results have not been positive. For rwo
> modestly successful examples, you can refer to the work of Alex Fraser
> (his "LEAF" model, ~2007 ACL) and Ulf Hermjakob (2009 EMNLP,
> specifically looking at improving Arabic-English alignment). Fazil
> Ayan wrote a series of papers on word alignment and tried to come up
> with evaluation criteria for predicting when they would help MT, but
> the whole enterprise has not been very successful (and, in my opinion,
> not a productive way to address the issue of improving MT translation
> models).
>
> Chris
>
> On Wed, Nov 4, 2009 at 2:58 PM, Mark Fishel <[email protected]> wrote:
>
>> Thanks for the responses, but that is not my question. All the three
>> mentioned papers evaluate the alignments either in comparison to a
>> reference alignment or in context of word-based SMT. What I am asking
>> is whether there's any work comparing the word alignment performance
>> in the phrase-based SMT pipeline. I just compared the default Moses
>> setup to my own implementation of IBM-1 (no reordering, no word
>> classes, etc.) substituting the 1st and 2nd steps (preparation and
>> GIZA++ word alignment parameter estimation) of
>> train-factored-phrase-model.perl: that resulted in only a small drop
>> of the BLEU score. It's only one test and one (relatively small)
>> corpus, so we can't draw conclusions from just that, but is there any
>> similar work?
>>
>> Thanks in advance,
>> Mark
>>
>>
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support
