You might consider cleaning your training working folder and starting
again. There's at least one report of (M)GIZA++ succeeding with one pass
of a corpus and then reporting similar hillclimb errors on a second run
with the exact same corpus.
http://article.gmane.org/gmane.comp.nlp.moses.user/1207/match=giza+hillclimb
I have experienced other hillclimb errors with corpora that have
extremely repetitive tokens between segment pairs or unnatural
machine-generated corpora. You might want to check your corpus for
things like "- - - - - - - - - - - - - -"
On 04/15/2013 06:58 PM, Hieu Hoang wrote:
You're unlikely to get a response to difficult questions on
giza++/mgiza these days. IMO, the knowledge about them is being slowly
lost as people who know about them (Och, Qin Gao, Abby Levenberg)
aren't available to answer these questions.
If you find the answer yourself, please let the mailing list know for
the future.
On 13/04/2013 14:34, Hassan Sajjad wrote:
Hi,
I am using mgiza for alignment. At model 4 1st iteration, hillclimb
gives the following warning
WARNING: already 47 iterations in hillclimb: 1.00001 0 79 48
*WARNING: DIFFERENT SUMS: (1) (-nan)*
WARNING: already 41 iterations in hillclimb: 3.28489 1 24 37
and after that p0_count is -nan which makes everything goes to -nan
later on.
The thing which I am doing different from a normal run of mgiza is
that I am restarting it after every iteration but changing t-table
probabilities. On a few parallel corpora, everything works fine.
Here is the link to the log file.
https://www.dropbox.com/s/e8z1l9w2i2hum1z/logError.txt
Any help would be highly appreciated.
Thanks.
Regards,
Hassan
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support
