This page shows how a trigram model for transfer ranking can be built. The first part shows how profiles with MRSs can be built and how MRS triples can be exported. It overlaps with DelphinTools and RedwoodsTop. The second part shows ErikVelldal's procedure for creating an MRS trigram model.

Contents

1. Objective
2. Preliminaries
3. Exporting MRS triples from Logon
   1. Creating the profiles
   2. Exporting triples from the profiles
4. Creating the transfer model
   1. Remove formatting
   2. Produce context cues
   3. Extract the vocabulary
   4. Train the model
5. Loading the transfer model

The procedure requires three tools:

• The CMU Toolkit (http://mi.eng.cam.ac.uk/~prc14/toolkit.html),

• SMT QuickRun (http://ufal.ms.mff.cuni.cz/~curin/SMT_QuickRun/), and

• DelphinTools.

The installation of the CMU Toolkit may conflict with Logon, so you may need to comment out the Logon settings in the .bashrc file temporarily while the installation is going on.

First, set a variable and a path:

TSDBHOME=$LOGONROOT/lingo/lkb/src/tsdb/home
export PATH=$LOGONROOT/lingo/lkb/src/tsdb/home:$PATH

This step is needed if you do not have a profile. It shows how you can get a profile from the object sentences of a bilingual corpus file. (See also DelphinTools.) Note that there are various other ways to do this. The first command below creates a new version of the bilingual corpus, where the object language comes first. The second command parses the sentences in the object language and stores the MRSs (of the five top ranked parses) in 'bitxt.'

logon_do --bitext PATH/TO/BILINGUAL/CORPUS/FILE $TSDBHOME/bitxt
logon_do --count 5 --limit 5 --grammar jaen --task omrs $TSDBHOME/bitxt

This command extracts triples from the top ranked MRSs of the profile created above. (See RedwoodsTop.)

$TSDBHOME/export --binary  --condition "result-id=0" --format triples bitxt/omrs

This is ErikVelldal's procedure for creating a transfer model, and his comments are given below. (I (PetterHaugereid) have added a couple of paths.)

First we remove all formatting inserted by the export code (to get only the tuples) and cat everything to a single file. Note that, in the pipe below, the script from the SMT_QuickRun package only inserts the "context cues" used by the CMU SLM toolkit, ie. the sentence boundaries <s> and </s>. (Note also that the path to the extracted triples '$LOGONROOT/tmp/bitxt.omrs/' may differ if you did not follow the procedure above.)

export PATH=/PATH/TO/SMT_QUICKRUN/bin:$PATH
find $LOGONROOT/tmp/bitxt.omrs/ -name *.gz | xargs zcat | awk '!/(^[\;\{\}\[]|^[[:space:]]*$)/' | add_sent_marks.prl | gzip > /tmp/mrstuples.gz

Produce a file holding the context cues, to be referenced by the CMU toolkit.

export PATH=/PATH/TO/CMU_TOOLKIT/bin:$PATH
echo "<s>" > ccs; echo "</s>" >> ccs

zcat /tmp/mrstuples.gz | text2wfreq > mrs.wfreq; cat mrs.wfreq | wfreq2vocab -top 65535 > mrs.vocab

zcat /tmp/mrstuples.gz | text2idngram -temp /tmp/ -n 3 -vocab mrs.vocab > mrs.idngram; idngram2lm -idngram mrs.idngram -n 3 -vocab mrs.vocab -binary mrs.binlm -calc_mem -context ccs -witten_bell

The MRS trigram model is written into the file 'mrs.binlm.'

The transfer model is loaded in an MT system's lkb/mt.lsp.

For example, if the transfer model is in the same directory:

(setf *transfer-tm*
 (namestring (make-pathname
              :directory (this-directory)
              :name "mrs.binlm")))