Hello Andy,
Date: Thu, 08 Jul 2004 14:28:12 +0100
From: Andy Way <[EMAIL PROTECTED]>
2. Can anyone envisage a situation where an SMT paper was asked to
compare its results against an MT model?
More than most other approaches, SMT people tend to ignore previous
work in the mistaken belief that it is not relevant, because SMT is
such a new paradigm. That is simply wrong, of course, as people have
argued (even in this mailing list). SMT is re-treading the path of
older approaches, but now doing things automatically that used to be
done by hand:
- the initial IBM work recreated word-replacement MT, but learned the
replacement rules automatically
- Och's and other current SMT is redoing EBMT, but learning the phrases
(i.e., examples) automatically
- Yamada and Knight, Wu, and Melamed each are working on versions of
transfer, with the rules, again, being learned
1. Can papers on EBMT succeed in getting published (especially in
non-expert, i.e. MT-specific, conferences) without making direct
comparisons to SMT?
Given the above trend, I think an effective response is to explicitly
say in an EBMT paper "yes I am doing EBMT but creating the example
phrases and their translation by hand; some SMT is creating the
phrases by machine; for me an open question is not only how to create
lots of patterns automatically but how good the actual patterns are",
which simultaneously shows familiarity with the relevant SMT work,
brings it into the picture in the right way, and addresses a point on
which SMT-style EBMT is vulnerable.
the bigger point, though, is: why should one not make comparisons to
SMT-style EBMT? A serious weakness of EBMT has always been the
bottleneck of building the example patterns and their translations
manually. SMT-style EBMT claims to overcome this bottleneck. Good
science demands that old-style EBMT work address this. You can still
then redirect the issue to the particular other, non-building, point
you are investigating.
3. Has EBMT as a paradigm been 'muscled out' by the more dominant
SMT approach?
I don't think so, because SMT is no longer a single unitary approach,
and will continue to split into flavors, just as MT has done. And
these flavors will increasingly correspond to the old approaches of
the Vauquois pyramid. SMT just does much of the tedious work
automatically, and in some cases much better, than humans can. EBMT
should not try to compete with SMT on its strength (brute force
learning of large sets of example patterns), but should co-opt SMT as
a technique to do its gruntwork.
The open questions of MT are still open: no-one can properly handle
interpersonal/stylistic/pragmatic effects of communication. If one
could use EBMT methods to capture style, for example, you'd be doing
something that syntax-based transfer approaches would find very hard
to do, and SMT approaches would struggle with given their need for
large corpora of unitary style.
E
--
Eduard Hovy
email: [EMAIL PROTECTED] USC Information Sciences Institute
tel: 310-448-8731 4676 Admiralty Way
fax: 310-823-6714 Marina del Rey, CA 90292-6695
http://www.isi.edu/natural-language/nlp-at-isi.html
_______________________________________________
MT-List mailing list
[EMAIL PROTECTED]
http://www.computing.dcu.ie/mailman/listinfo/mt-list