A few more thoughts from the Workshop on Parallel Text. One of the
interesting points that was raised during the conference in general,
and in particular during this workshop, was that there does not seem
to be a correlation between Alignment Error Rates (AER) and Machine
Translation "accuracy" or quality. Alignment is the process of deciding
which words and phrases are translations of each other in a parallel
text, so it seems a bit count-intuitive (to me) that there would not
be a connection between AER and translation quality. But, this seems to
be the consensus view among people who work in MT, suggesting perhaps
that current methods are very robust to errors in the aligned parallel
text that is used to train translation models. I can't imagine that
MT output would be very good at all if the alignments in the paralllel
text were totally wrong.
Another point which arose was (during Franz Och's talk) that current
MT methodologies don't really use syntactic or semantic information. In
Franz's opinion, word sense disambiguation does not enter into current
MT models as an explicit component, since it is already (implicitly)
being taken care of. This is an interesting notion, and I think I
understand it. In current MT models, phrases take on an important role.
In other words, you are not translating word by word, but rather phrase
by phrase, in the best case at least. Well, phrases are not nearly
as ambiguous as are single words. For example, if you look at WordNet
compounds (concepts with two or more words, such as winston_churchill or
racing_car) in general the level of ambiguity that those compounds have
is much less than those of single words. So perhaps that is how the
ambiguity is being dealt with. There is also the issue that the act of
alignment will help to resolve some ambiguities (translational). Some
words that are ambiguous in one language will translate to completely
different forms depending on the sense in which they are used in the
source language, so while there may be ambiguity in one language, the
alignment to the target language essentially resolves that ambiguity.
So I think it might be that a combination of phrase based methods along
with exploiting translational equivalence might well help reduce the
impact of ambiguity on MT.
There was also a paper at ACL 2005 that quantifies and studies these
issues more formally, and again suggests that WSD does not help MT. I
think this is a conclusion I had arrive informally (based on the reasoning
above) but this paper does a very nice job of making such arguments
more rigorous and convincing.
@InProceedings{carpuat-wu:2005:ACL,
author = {Carpuat, Marine and Wu, Dekai},
title = {Word Sense Disambiguation vs. Statistical Machine
Translation},
booktitle = {Proceedings of the 43rd Annual Meeting of the Association for
Computational Linguistics (ACL'05)},
month = {June},
year = {2005},
address = {Ann Arbor, Michigan},
publisher = {Association for Computational Linguistics},
pages = {387--394},
url = {http://www.aclweb.org/anthology/P/P05/P05-1048}
}
--
Ted Pedersen
http://www.d.umn.edu/~tpederse
Yahoo! Groups Links
<*> To visit your group on the web, go to:
http://groups.yahoo.com/group/nlpatumd/
<*> To unsubscribe from this group, send an email to:
[EMAIL PROTECTED]
<*> Your use of Yahoo! Groups is subject to:
http://docs.yahoo.com/info/terms/
