Cross today's paper with what we did on Nov 3rd, 2003.


================================================================================
	
From Kevin Duh:

1) The issue of imposing reordering constraints is interesting if seen
from a bias-variance tradeoff perspective. Both the IBM reordering
constraints and the ITG constraints limit the complexity of the models
in some way. I'm interested to see if anyone has done deeper analysis
on exactly what kinds of reorderings IBM constraints (or ITG
constraints) can and cannot cover for a set of language pairs. This
may be more insightful than simply comparing the BLEU and NIST scores
of IBM vs ITG for just one language pair.

So far my only hunch is that "Subject-Verb-Object and
Subject-Object-Verb language pairs have long range reorderings, which
may imply that ITG-type constraint is better suited."  Beyond this, I
have no idea why one constraint would work better than the other. So
it would be interesting to explore this; it'll not only help with MT
performance but also shed light on linguistics research as well.

2) I wonder if ITG constraints will be more useful if the phrases used
are linguistically-meaningful phrases rather than automatically
generated phrases. In paraphrase examples like "Let's meet at the
fountain tomorrow morning" vs "Let's meet tomorrow morning at the
fountain", it seems that the ITG reordering of "chunks" correspond to
linguistically-meaningful phrases (e.g. "at the fountain", "tomorrow
morning"). But when we have automatically-generated phrases like
"fountain tomorrow", then it seems a lot harder to do ITG reordering.

3) What's the difference between applying reordering constraints at
the word level vs. at the phrase level? The authors of this paper also
wrote a paper comparing IBM and ITG for word MT systems (ACL 2003). I
wonder if anything insightful can be gleaned from comparing the two
results.

================================================================================