$Date: 2004/02/09 08:11:23 $

<hr>

From Kevin Duh:
<p>
Random thoughts for the notes:
<p>
It's interesting how we swing back and forth between more constrained
and less constrained models for MT. Yamada and Knight proposed a
syntax-based TM with reorder operation, etc, which captures the
groupings of words in language translation. Then, we realize that a
simple tree is too constrained, so Gildea proposes a "loosely
tree-based" model, which uses grouping and cloning to get around the
fact that the Yamada and Knight model cannot get some word orderings.
<p>
I wonder if there exists some "optimal" model with just the right
amount of contraints and parameters to be statistically learned. I'm
starting to think that different models will perform better for
different pairs of languages. If so, it implies the design of
statistical models (be it IBM, tree-based, phrase-based, loosely-tree
based) should still be motivated largely by linguistic knowledge.
<p>

<hr>

Bilmes: What is the inherent advantage of this technique vs.
considering multiple probabilistic parses, rather than using one fixed
parse that gets manipulated via a set of operations, some of which
might make the parse ungrammatical. Should such approaches be
compared. Again, this seems to be a case of ignoring "uncertainty"
that might potentially be useful.
<p>
Also, who is Alexander Calder? He invented the mobile,
see <a href="http://www.enchantedlearning.com/artists/calder/">here</a>

<hr>

Jeremy Kahn:
A recent LanguageLog post illuminates interesting problems with
segmentation of Japanese patent language for MT to English:
<a href="http://itre.cis.upenn.edu/~myl/languagelog/archives/000432.html">
link</a>
<p>
I won't give away the punchline but it boils down to a tokenization
problem.
<p>
Seems to be my week for thinking about tokenization problems.
<hr>