hi boxing
i had a look at this and there is a bug in the code.
the alignment output doesn´t take into account the fact that hypotheses with
different target lengths can be combined during stack pruning. apologies
will check in the fix in the next few days, as soon as i can connect my
laptop to the net. check the svn commits for when
http://sourceforge.net/mailarchive/forum.php?forum_name=mosesdecoder-commits
Boxing Chen <[EMAIL PROTECTED]> wrote:
Dear moses-developer,
Currently, I am using moses of version 2008-02-20.
if I use the option: -include-alignment-in-n-best to keep the phrase alignment
informarion in the nbest list, more than half hypotheses got wrong phrase
alignments, such as:
0 ||| oh , the flight is the c 3 0 6 . ||| d: 0 -3.77112 0 0 -2.52851 0 0 lm:
-54.4133 tm: -8.65867 -10.4881 -6.35404 -7.50889 5.99938 w: -11 ||| -2.98287
||| 0-1=0-1 2=2 3-4=3-4 5=5-6 6=7 7-9=7-9
in this example, the eighth target word ("3") is aligned twice by: 6=7 7-9=7-9
3 ||| 2 passengers is . ||| d: 0 -2.81628 0 0 -1.6697 0 0 lm: -27.6149 tm:
-23.9911 -20.0994 -7.60809 -5.73339 3.99959 w: -4 ||| -3.15759 ||| 0-1=0 2-3=1
4=4 5-6=5
there are only 4 target words, but the alignment indicates to the fifth and
sixth target words: 4=4 5-6=5
in total, out of 2,440,820 hypotheses (IWSLT'06 dev set, 5000-best), there are
1,780,673 got wrong phrase alignment.
best regards,
-Boxing
____________________________________________________________________________________
Be a better friend, newshound, and
know-it-all with Yahoo! Mobile. Try it now.
http://mobile.yahoo.com/;_ylt=Ahu06i62sR8HDtDypao8Wcj9tAcJ
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support
Hieu Hoang
http//www.hoang.co.uk/hieu
---------------------------------
Sent from Yahoo! Mail.
A Smarter Email._______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support
