Automatically trimming a monodix

There used to be a problem in Apertium regarding copies. When we came to make a new language pair based off some existing resource in Apertium (e.g. a monodix), we would make a copy of that resource, and then change it as we need. This was less than ideal because:

• it meant that any improvements we made weren't automatically carried over to the dictionary we copied from
• it meant that any improvements in the dictionary we copied from weren't carried over into our new dictionary.

So why did we do it ? -- Testvoc. If we had entries in our monodix that were not in our bidix, then we got lots of @ in our output. This is bad.

One way we got around the problem was to use ad hoc scripts for trimming the dictionaries (see for example: apertium-af-nl, apertium-sme-nob and the trim-lexc.py script in trunk/apertium-tools), but these were less than ideal because usually they had to include specific hacks for differences in dictionary format.

The new and improved way is to take the intersection of our monodix and our bidix binaries, and use that for analysis. That's what is described below.

Note: This doesn't take care of the whole testvoc problem. It is still necessary to get rid of # symbols, which often arise due to bugs in transfer code.

Example

Suppose you have the monodix:

<dictionary>
  <alphabet>abcdefghijklmnopqrstuvwxyz</alphabet>
  <sdefs>
    <sdef n="n"/>
    <sdef n="sg"/>
    <sdef n="pl"/>
  </sdefs>
  <pardefs>
    <pardef n="beer__n">
      <e><p><l></l><r><s n="n"/><s n="sg"/></r></p></e>
      <e><p><l>s</l><r><s n="n"/><s n="pl"/></r></p></e>
    </pardef>
  </pardefs>
  <section id="main" type="standard">
    <e lm="beer"><i>beer</i><par n="beer__n"/></e>
    <e lm="school"><i>school</i><par n="beer__n"/></e>
    <e lm="computer"><i>computer</i><par n="beer__n"/></e>
    <e lm="house"><i>house</i><par n="beer__n"/></e>
  </section>
</dictionary>

It generates the following strings:

$ lt-expand test-en.dix
beer:beer<n><sg>
beers:beer<n><pl>
school:school<n><sg>
schools:school<n><pl>
computer:computer<n><sg>
computers:computer<n><pl>
house:house<n><sg>
houses:house<n><pl>

But our bidix is only:

<dictionary>
  <alphabet>abcdefghijklmnopqrstuvwxyz</alphabet>
  <sdefs>
    <sdef n="n"/>
    <sdef n="sg"/>
    <sdef n="pl"/>
  </sdefs>
  <pardefs>
  </pardefs>
  <section id="main" type="standard">
    <e><p><l>beer<s n="n"/></l><r>garagardo<s n="n"/></r></p></e>
    <e><p><l>house<s n="n"/></l><r>etxe<s n="n"/></r></p></e>
  </section>
</dictionary>

We don't want to include the entries for "computer" and "school", because then we would get @ in our output.

HFST

Here is a Makefile that given two dictionaries test-en.dix (the monodix) and test-en-eu.dix (the bidix), will produce a binary transducer of the monodix in HFST format, that only contains the strings matching prefixes in the bidix.

all:
	lt-comp lr test-en.dix test-en.bin
	lt-comp lr test-en-eu.dix test-en-eu.bin
	lt-print test-en.bin > test-en.att
	lt-print test-en-eu.bin > test-en-eu.att
	hfst-txt2fst -e ε <  test-en.att > test-en.fst
	hfst-txt2fst -e ε <  test-en-eu.att > test-en-eu.fst
	hfst-invert test-en.fst -o test-en.mor.fst
	hfst-project -p upper test-en-eu.fst > test-en-eu.en.fst
	echo " ?* " | hfst-regexp2fst > any.fst
	hfst-concatenate -1 test-en-eu.en.fst -2 any.fst -o test-en-eu.en-prefixes.fst
	hfst-compose-intersect -1 test-en-eu.en-prefixes.fst -2 test-en.mor.fst | hfst-invert -o test-en.trimmed.fst


clean:
	rm *.bin *.att *.fst

If we run hfst-fst2strings, we get:

$ hfst-fst2strings test-en.trimmed.fst
beer:beer<n><sg>
beers:beer<n><pl>
house:house<n><sg>
houses:house<n><pl>

Compounds vs trimming in HFST

The sme.lexc can't be trimmed using the simple HFST trick, due to compounds.

Say you have cake n sg, cake n pl, beer n pl and beer n sg in monodix, while bidix has beer n and wine n. The HFST method without compounding is to intersect (cake|beer) n (sg|pl) with (beer|wine) n .* to get beer n (sg|pl).

But HFST represents compounding as a transition from the end of the singular noun to the beginning of the (noun) transducer, so a compounding HFST actually looks like

((cake|beer) n sg)*(cake|beer) n (sg|pl)

The intersection of this with

(beer|wine) n .*

is

(beer n sg)*(cake|beer) n (sg|pl) | beer n pl

when it should have been

(beer n sg)*(beer n (sg|pl)

Lttoolbox doesn't represent compounding by extra circular transitions, but instead by a special restart symbol interpreted while analysing, so lt-trim is able to understand compounds by simply skipping the compound tags.

lttoolbox actually has another type of multiword that is more similar to how compounds in hfst work: the <j/> (JOIN) multiword. Here the full path is in the FST like foo<tag>+bar<othertag>. This is handled by lt-trim: When it sees a +, it moves to the start in bidix but keeps going from where it was in monodix.

In sme-nob, we ensured that anything that we wanted pretransfer to split would have a + in it, by using tag relabelling. So the FST would have paths like foo<tag>+bar<othertag> which are split into foo<tag> and bar<othertag>. (Unlike with JOIN, we can have several +'es in a row.) Perhaps a special-purpose hfst trimming tool could do what lt-trim does on seeing the plus.

The downside is that lexc writers have to really ensure that anything that is supposed to be as two units in bidix really has a +, but they have to do that anyway!

lttoolbox

The command is named lt-trim.

$ lt-comp rl apertium-en-ca.en-ca.dix ca-en.autobil.bin

$ lt-comp lr apertium-en-ca.ca.dix ca-en.automorf-full.bin 

$ lt-trim ca-en.automorf-full.bin  ca-en.autobil.bin  ca-en.automorf.bin

where the left side of the second transducer (ca-en.autobil.bin) is altered so it

• gets a ".*" appended to it (so if "foo<vblex>" is in there, it will let through "foo<vblex><pres>"), and then so it
• has all lemq's moved to be after the tags (so if "foo<vblex># bar" is in there, it will let through "foo<vblex><pres># bar"),

and only strings from the first transducer whose right side (analysis) match the input side of the altered second transducer are included into the final compilation.

See also

• Testvoc
• Why we trim