Dictionary coverage

From Apertium
Revision as of 18:07, 3 November 2009 by Jacob Nordfalk (talk | contribs)
Jump to navigation Jump to search

There exists an experimental tool in in Apertium-dixtools to do frequency statistics on a dictionary, i.a. which entries are used and which entries are not used.

This is i.a. usable if have a language pair in which you want to do the opposite direction. The tool was successfully on English and Esperanto to make the eo-en direction. Please contact me (--Jacob Nordfalk 18:07, 3 November 2009 (UTC)) for help using it.


Here is how it could be applied on sv-da to analyse how much of the Danish dictionary (which contains many entries not used from sv to da).


First remove duplicates in dixes, e.g. do:

 apertium-dixtools fix -alignBidix apertium-sv-da.sv-da.dix apertium-sv-da.sv-da.dix
 apertium-dixtools fix -alignMonodix apertium-sv-da.da.dix apertium-sv-da.da.dix
 apertium-dixtools fix -alignMonodix apertium-sv-da.sv.dix apertium-sv-da.sv.dix

Then, make a copy of the pair in the profiler/ subdirectory.

 $ mkdir profiler/
 $ cp * profiler/*


Then, create a "profiler" version of your dictionaries:

 $ apertium-dixtools profilecreate . sv-da 

This will create a file dixtools-profilekeys.txt and overwrite the profiler/*.dix dictionaries. Now do

 $ cd profiler
 $ make

Then edit the modes.xml and add a mode where you replace

     <program name="lt-proc $1">
       <file name="sv-da.autogen.bin"/>
     </program>

with usage of the profiling version of the dictionary, after which the task apertium-dixtools profilecollect must be inserted to collect the data (saving it to dixtools-profiledata.txt) and filter the output for the following stages in the mode file:

     <program name="lt-proc $1">
       <file name="profiler/sv-da.autogen.bin"/>
     </program>
     <program name="apertium-dixtools profilecollect">
       <file name="dixtools-profiledata.txt"/>
     </program>

Use the mode, for example on a corpus and/or with your favorite testvoc script:

 $ apertium-dixtools profileresult
 Reading dixtools-profilekeys.txt
 Reading dixtools-profiledata.txt
 Writing dixtools-profileresult.txt


How it works

It is adding keys to the dixes. So

<pardef n="b/urde__vbmod">
  <e>       <p><l>urde</l>      <r>urde<s n="vbmod"/><s n="inf"/><s n="actv"/></r></p></e>
  <e>       <p><l>ør</l>        <r>urde<s n="vbmod"/><s n="pres"/><s n="actv"/></r></p></e>
  <e>       <p><l>urde</l>      <r>urde<s n="vbmod"/><s n="past"/><s n="actv"/></r></p></e>
  <e>       <p><l>urdet</l>     <r>urde<s n="vbmod"/><s n="pp"/></r></p></e>
</pardef>

<pardef n="må/tte__vaux">
  <e>       <p><l>tte</l>       <r>tte<s n="vaux"/><s n="inf"/><s n="actv"/></r></p></e>
  <e>       <p><l></l>          <r>tte<s n="vaux"/><s n="pres"/><s n="actv"/></r></p></e>
  <e>       <p><l>tte</l>       <r>tte<s n="vaux"/><s n="past"/><s n="actv"/></r></p></e>
  <e>       <p><l>ttet</l>      <r>tte<s n="vaux"/><s n="pp"/></r></p></e>
</pardef>

becomes

<pardef n="b/urde__vbmod">
  <e>       <p><l>urde</l>      <r>urde<s n="vbmod"/><s n="inf"/><s n="actv"/></r></p><p><l>%bct%</l><r/></p></e>
  <e>       <p><l>ør</l>        <r>urde<s n="vbmod"/><s n="pres"/><s n="actv"/></r></p><p><l>%bcu%</l><r/></p></e>
  <e>       <p><l>urde</l>      <r>urde<s n="vbmod"/><s n="past"/><s n="actv"/></r></p><p><l>%bcv%</l><r/></p></e>
  <e>       <p><l>urdet</l>     <r>urde<s n="vbmod"/><s n="pp"/></r></p><p><l>%bcw%</l><r/></p></e>
</pardef>

<pardef n="må/tte__vaux">
  <e>       <p><l>tte</l>       <r>tte<s n="vaux"/><s n="inf"/><s n="actv"/></r></p><p><l>%bcx%</l><r/></p></e>
  <e>       <p><l></l>          <r>tte<s n="vaux"/><s n="pres"/><s n="actv"/></r></p><p><l>%bcy%</l><r/></p></e>
  <e>       <p><l>tte</l>       <r>tte<s n="vaux"/><s n="past"/><s n="actv"/></r></p><p><l>%bcz%</l><r/></p></e>
  <e>       <p><l>ttet</l>      <r>tte<s n="vaux"/><s n="pp"/></r></p><p><l>%bd0%</l><r/></p></e>
</pardef>

The dixtools-profileresult.txt contains

0 bct <e><l>urde</l><r>urde<vbmod><inf><actv></r></e>
0 bcu <e><l>ør</l><r>urde<vbmod><pres><actv></r></e>
0 bcv <e><l>urde</l><r>urde<vbmod><past><actv></r></e>
0 bcw <e><l>urdet</l><r>urde<vbmod><pp></r></e>
1 bcx <e><l>tte</l><r>tte<vaux><inf><actv></r></e>
41 bcy <e><l></l><r>tte<vaux><pres><actv></r></e>
1 bcz <e><l>tte</l><r>tte<vaux><past><actv></r></e>
1 bd0 <e><l>ttet</l><r>tte<vaux><pp></r></e>

so you can see that the first paradigm isnt used at all and in the 2nd paradigm the %bcy% was used 42 times and the other entries 1 time each.