Re: [Mt-list] Where is MT at today?

[Michel Simard]

Hello all,
Coming back to Andy's questions:
   1. Can papers on EBMT succeed in getting published (especially in
      non-expert, i.e. MT-specific, conferences) without making direct
      comparisons to SMT?
Certainly one giant step in this direction would be made if people doing
SMT of the phrase-based flavor (i.e. just about everybody in SMT these
days) acknowledged the fact that, as Harry and Ed rightly point out,
what they are doing is really just EBMT with a statistical twist.
Unfortunately, in my experience, SMT people seem very reluctant to admit
this.  Why this is so is not so clear, but I suspect that the EBMT label
has a bit of a dubious reputation among the machine-learning crowd.  And
maybe one  of the reasons for that is that very few EBMT papers report
on quantitative evaluations.
While subjective human MT evaluations are very costly, there are now a
number of alternatives (BLEU, NIST, recall/precision) which are cheap,
easy to use, and available for anyone with an internet connection.  And
in spite of all the controversy that these methods stir, they are better
than nothing.
   2. Can anyone envisage a situation where an SMT paper was asked to
      compare its results against an MT model?
I think an SMT paper that doesn't report on empirical comparisons with 
competing systems or approaches is just as likely to get rejected as 
Andy's.   For better or for worse, that's the name of the game these 
days. But you can't empirically compare against approaches for which 
there are no empirical evaluations.  So SMT papers report comparisons 
with what's available: other SMT systems.

Cheers,
Michel


_______________________________________________
MT-List mailing list
[EMAIL PROTECTED]
http://www.computing.dcu.ie/mailman/listinfo/mt-list