Hello everyone,
thanks for your kind advises and tips! Sadly i can't use any tool other
than GIZA++ since the post-processing stage, pisa, needs GIZAs parameter
dump files. Also, what i try to do is not grapheme<->phoneme alignment,
but word-in-long-utterance<->phoneme-string alignment. Pisa was
originally made for a much harder problem, words in one language to
phoneme-strings in another language, but since it performs so well on
that problem we decided to use it for our problem too. The initial
alignments which still include the distortion step look very promising,
almost all errors that can be seen look like they are due to the
non-monotonity of the alignment. So we set out to try to get rid of the
distortion step. The raw GIZA++ alignments for this problem were - by
the way - barely usable at all ;(. So, if you're looking at a similar
problem i can warmly recommend PISA. I think theres also a paper upcoming.
But anyways, back to my problem:
On 08/24/2012 08:23 AM, Marcin Junczys-Dowmunt wrote:
> If the data is monotone then it rather improbable that Giza will
> produce crossing alignments. I'd guess it's just worth a try.
I've already tried this and the alignment is rather pulled apart. There
is barely one word (in a 115617 word, 663177 phoneme corpus) that is
actually monotone without a phoneme from another word mixed in or a
phoneme from this word stolen into another word. While i'm not capable
enough in SMT to analyze why this is, my feeling tells me that by simply
tuning a few parameters the fundamental problem here won't vanish?
(PLEASE, correct me if i'm wrong, i'm still a really bad rookie in this
field ;P)
On 08/24/2012 04:10 AM, Chris Dyer wrote:
> I think adding this would be have tremendous value.
Yay! I'm not alone ;P. Especially in conjunction with PISA, i think this
would open the door to a vast improvement of alignment quality for many
use-cases.
> It should be possible to adapt Giza's HMM implementation to produce
> monotone alignments. These are the changes that would be necessary
> (and which should be fairly easy, if you can figure out the code):
Unfortunately i'm still not familiar with the code. I'm reading and
reading, but the almost complete absence of comments and the rather
chaotic (?? to me at least ??) organization of the code doesn't make it
easy for me :/.
> 1) alignment distribution initialization. by default Giza initializes
> the HMM transition probabilities to be uniform (effectively making the
> first iteration of HMM training the same as one more iteration of
> Model 1). You would need to alter this to make "reverse" jumps have
> probability 0.
>
> 2) smoothing. by default, Giza does something to prevent probabilities
> from ending up zero (maybe add alpha?). This is fine for monotone
> jumps, but you want to make sure that "backward" jumps end up zero.
So, for these Tips i'm really grateful! I've already tried to track down
the "model5smoothfactor" parameter (since i assumed that would lead me
to where distortion happened, since its doc-tag says its the distortion
smoothing). Sadly i had to find out that it looks like this parameter
isn't used at all, since the parsing is commented out?!?
main.cpp:
//makeSetCommand("model5smoothfactor","0.0",getGlobalParSet(),2);
A few tests on a small corpus with values of 1000, 1, 0, -1, -1000
yielded *exactly* the same results! I'm a bit confused now ... is it to
be expected that some of the GIZA++ parameters aren't even implemented
despite the help advertising them?
Anyways, i'll try to dig around in the code a bit more based on your
tips. But, if you happen to be familiar with the code, i'd be really
really grateful for some even more detailed pointers :(. I have that bad
feeling that this will take a really really long time if i'll wander the
maze of GIZA code all alone ;P.
Thanks to you all for your help and support! And Best Regards from Germany!
- Dario Ernst
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support
