[developers] quick check/checkpaths

Francis Bond fcbond at gmail.com
Fri Oct 31 06:49:35 CET 2008


G'day,

Thanks  Berthold and Ann.


> On Thu, 2008-10-30 at 15:17 +0000, Ann Copestake wrote:
>> My guess is that even 45 is more than necessary - I seem to remember
>> that the vector stayed much the same after 5 short items when we tried
>> it originally in Verbmobil days.  BUT - it occurs to me - is the size
>> of the vector still set to 30 for PET?  It might be worth checking
>> that this is still around optimal if nobody has done that recently.
>>
>
> >From 40 qc paths on, the positive effect declines with GG, at least as
> measured for Pet a year or two ago.


I think I will just run on the MRS test set then (100 and a bit items).

Yours,

-- 
Francis Bond <http://www2.nict.go.jp/x/x161/en/member/bond/>
NICT Language Infrastructure Group

P.S.  This is how I now do it (from jacy/utils/make-qc.bash).

#!/bin/bash
# from the main grammar directory
unset DISPLAY;
unset LUI;

lkbdir=${LOGONROOT}/lingo/lkb;
grammardir=${LOGONROOT}/dfki/jacy;

### make input file
cut -d@ -f7 $grammardir/tsdb/skeletons/mrs/item  > $grammardir/utils/mrs.txt

#
# CHEAP
#

# back up quick check file
mv $grammardir/pet/qc.tdl $grammardir/pet/qc.tdl.old

#flop the grammar once
cd $grammardir
flop japanese

# calculate the quickcheck file
cat $grammardir/utils/mrs.txt | \
cheap -limit=50000 -packing -compute-qc=pet/qc.tdl $grammardir/japanese

# flop the grammar again
cd $grammardir
flop japanese

echo "PET done"
###
### LKB
###
# back up
mv $grammardir/lkb/checkpaths.lsp $grammardir/lkb/checkpaths.lsp.old

### FIXME should redo the input file at somestage

{
 cat 2>&1 <<- LISP
  (load "$lkbdir/src/general/loadup")
  (compile-system "lkb" :force t)
  (lkb::read-script-file-aux  "$grammardir/lkb/script")
  ;;; set an edge limit
  (setf  lkb::*maximum-number-of-edges* '5000)
  ;; make the checkpaths
  (lkb::with-check-path-list-collection
   "$grammardir/lkb/checkpaths.lsp"
    (lkb::parse-sentences
    "$grammardir/utils/mrs.txt"
    "/tmp/mrs.txt.out"))
  (format t "~%All Done!~%")
  #+allegro        (excl:exit)
  #+sbcl           (sb-ext:quit)
LISP
} | ${LOGONROOT}/bin/logon --source -I base -locale ja_JP.UTF-8

echo "LKB done"

echo "please commit the new files"
echo "svn commit -m 'new quick check paths' pet/qc.tdl lkb/checkpaths.lsp"



More information about the developers mailing list