I just wanted to chip in a couple of cents from my side.
First, I agree with Tom: my most sincere admiration for the patience you
all have evidenced when answering in this thread. It is certainly not easy
to keep calm when someone is saying that all your work in the last 10
years is wrong.
Second, I think it would be a really good idea to pin this thread
somewhere, perhaps clean it up a bit, and save it for the future: I think
future researchers and NLP students would be able to learn a real lot by
reading through it, it might constitute the most comprehensive guide to
explaining [at least some aspects of] the decoding process I have read so
far.
Finally, regarding James' claims:
1) So if I've understood you correctly you are saying we have a system
that is purposefully designed to perform poorly with a disabled LM and
this is the proof that the LM is the most fundamental part.
Nope. This misses the point. The system is designed to optimize something
that is unoptimized. It *does not care* about what happens when sticks are
thrown into the system's wheels. It's like a gas engine: it's designed to
do the most out of the gas you poor into it. It *does not care* about what
happens when you poor olive oil into it. The point is not about
purposefully designing something to perform poorly in poor conditions, the
point is about designing something that performs as good as possible with
the conditions we have: because, at the end of the day, what we care about
is having a system that is able to yield good (or fair) translation
quality.
2) If you still feel that the LM is the most fundamental component I
challenge you to disable the TM and perform LM only translations and see
what kind of BLEU scores you get.
I think Marcin meant that the LM is the strongest feature from all the
individual features available in the system. Disabling the TM is disabling
a bunch of individual features, not just one. And anyway, this misses the
point too: we care about doing the best we can with what we have, not
about what happens when you poor olive oil into a gas engine.
Best,
Germán Sanchis-Trilles
Co-founder Sciling S.L.
Tel: +34 658272899
email: gsanc...@sciling.com
_______________________________________________
Moses-support mailing list
Moses-support@mit.edu
http://mailman.mit.edu/mailman/listinfo/moses-support
