Subreadings in Constraint Grammar

This is now implemented in vislcg3: http://beta.visl.sdu.dk/cg3/chunked/subreadings.html

Why we need sub-readings

Typical input with sub-readings:

^foobar/foo+bar/fubar/flue+barge$

Right now, only the last sub-reading is used, in the above example, vislcg3 treats it as if it were

^foobar/bar/fubar/barge$

This works great for compounds where the stuff before the + is mostly inconsequential, while for other multiword expressions it is not so good... (Also, mapping tags are only put on the last sub-reading now.)

Wait can't we just split on the + with pretransfer before sending this to cg-proc?

No, because we first have to disambiguate between eg. ^foobar/foo+bar/fubar/flue+barge$ (what would that even look like if split? wouldn't work)

What we need

• We may need to refer to a non-main sub-reading in order to disambiguate
• We may want to put a mapping tag on a non-main sub-reading
• And of course we want to be able to refer to the main sub-reading

Referring to the final sub-reading

Northern Sámi postpositions take genitive.

Input fragment:

^soahtefámu/soahti<N><Sg><Nom><Cmp>+fápmu<N><Sg><Acc>/soahti<N><Sg><Nom><Cmp>+fápmu<N><Sg><Gen>$ 
^vuostá/vuostá<Po>/vuostá<Pr>/vuostá<N><Sg><Nom>$

Correct output:

^soahtefámu/soahti<N><Sg><Nom><Cmp>+fápmu<N><Sg><Gen><@→P>$        # war.power.GEN
^vuostá/vuostá<Po><@←ADVL>$^                                       # against.PO

If the input noun were unambiguously nominative, the Po reading should not be selected, so we might have a rule somewhere with

REMOVE Po if (-1 (Nom))

but if this matched non-final sub-readings, we would get the wrong tagging here. By default, non-final sub-readings are ignored, so the sme-nob CG's work fine (as do the nn-nb ones for compounding there).

Referring to non-final sub-readings

Input:

^D'an/Da<pr>+an<det><def><sp>$
^emgann/emgann<n><m><sg>$ 
^ez/e<vpart><obj>/ael<n><m><pl>/mont<vblex><pri><p2><sg>/monet<vblex><pri><p2><sg>/e<pr>+da<det><pos><mf><sp>$
^an/an<det><def><sp>/mont<vblex><pri><p1><sg>/monet<vblex><pri><p1><sg>$

Correct output:

^D'an/Da<pr><@ADVL→>+an<det><def><sp><@→N>$       # to.the
^emgann/emgann<n><m><sg><@P←>$                    # battle
^ez/e<vpart><obj><@Pcle>$                         # PART
^an/mont<vblex><pri><p1><sg><@+FMAINV>$           # I.go

• We want to refer to the <pr> sub-reading when mapping emgann as @P← (possibly also in disambiguation).
• We want to MAP an @ADVL→ tag on the <pr> sub-reading (also a @→N tag on the determiner). These sub-readings are split into two units by pretransfer.

VISL CG-3 syntax

VISL CG-3 keeps the default behaviour that we always refer to only the last sub-reading unless explicitly mentioning sub-readings. But for some languages, you might want to prefer the first sub-reading to be main by default. VISL CG-3 caters to both preferences. From the manual:

The order of which is the primary reading vs. sub-readings depends on the grammar SUBREADINGS setting:

     SUBREADINGS = RTL ; # Default, right-to-left
     SUBREADINGS = LTR ; # Alternate, left-to-right

Then, to refer to a non-final sub-reading in the default RTL mode, we could say

 ADD (@ADV←) TARGET (n) IF (-2/1 (pr)) (-2 (n)) ;

to say that we require the next-to-final sub-reading of the cohort two positions left be a word that has the main reading n and next sub-reading pr. This would match if the input were e.g.

 ^forsooth/for<pr>+sooth<n>/forsooth<adv>$ ^he/prpers<prn>$ ^be/be<vblex>$

Since we only have two sub-readings here, we could also ask that the last sub-reading be pr, with the same effect:

 ADD (@ADV←) TARGET (n) IF (-2/-1 (pr)) (-2 (n)) ;

Parallell to regular CG word indexes, 0 is the "head". In RTL mode, this is the last sub-reading, while -1 is one sub-reading to the left of that. Positive numbers read from the left, so 1 is the first sub-reading from the left. For three sub-readings, that gives us the following indexing:

   ^foo<tags>+bar<tags>+fie<tags>$
      2        1         0
     -1       -2        -3

For LTR mode, the left sub-reading is the head with index 0, and counts go the other way:

   ^foo<tags>+bar<tags>+fie<tags>$
      0        1         2
     -3       -2        -1

To ADD the tag to the non-final sub-reading itself, use the SUB:N keyword after ADD:

 ADD SUB:-1 (@→V) TARGET (pr) IF (*1 (v)) ;

We might also want to say "require any main- or sub-reading to be tagged pr":

 ADD (@P←) TARGET (n) IF (-1/* (pr)) ;

or to say that all readings of the previous word are unambiguously pr (on one of the sub-readings):

 ADD (@P←) TARGET (n) IF (-1C/* (pr)) ;

You can now do

REMOVE SUB:* (pr) IF (1 (vblex));

to remove any reading which has a pr on some sub-reading if there's a following verb.

The old workaround was to do

REMOVE (pr) IF (1 (vblex));
REMOVE SUB:1 (pr) IF (1 (vblex));
REMOVE SUB:2 (pr) IF (1 (vblex));
REMOVE SUB:3 (pr) IF (1 (vblex));

etc. as high as your analyser allowed.

Wishlist

A special set for "has-subreading"

The same way that you can do

LIST match-any = (*)

it would be nice to be able to do e.g.

LIST compound = (*/1)

to match on any (sub)reading that has at least one subreading.

Note: This is typically meant to be used on the main reading, so that you can have rules like

REMOVE compound IF …

or

MAP (@FOO) IF (-1 compound + Pl LINK …)

(note that the (*/1) is a "tag" of the main reading here, so the + Pl means that the main reading is plural, and there is at least one subreading below the main reading).

(a)+SUB:1(b) – requirements on both main and sub-readings at once

There is currently no way for SELECT/REMOVE to target tags in both the sub-reading and the reading. E.g. say you use the default mode where final readings are main, and your input is

^foobar/foo<vblex>+bar<n>/foo<n>+bar<n>/foo<vblex>+bar<vblex>/foo<n>+bar<vblex>$

Now given a verb-verb compund is less likely than an anything-noun compound, you want to remove any reading where both sub-readings are verbs. If you try

REMOVE SUB:1 (vblex) (0 (vblex)) ;

it will wrongly remove the verb-noun reading as well, since the (0 (vblex)) matches on another reading's sub-reading. If you try

REMOVE (vblex) (0/1 (vblex)) ;

it will wrongly remove the noun-verb as well, since the (0/1 (vblex)) matches on another reading's sub-reading.

Similarly, there is no way to MAP/ADD tags to the main reading of a reading that has a certain sub-reading. Using the above example, there is no way to map @v to only the verb-verb compound without either hitting the noun-verb or the verb-noun as well.

Possible syntax, similar to set intersection:

REMOVE (vblex) + SUB:1 (vblex);

This might make sense inside a context condition as well:

REMOVE (vblex) IF (0 (n) + 0/1 (n));

Or even as a variable, assuming people don't name their sets "SUB:1":

SET verb-verb-compound = (vblex) + SUB:1 (vblex);