Use correct spell-relax file about plains-cree-fsts HOT 7 CLOSED

Depends with what script you compose the descriptive and normative FSTs.

If you compose it with the normative analyzer as done in the crk.hfscript recipe in crk/inc (as you plonk it to the input side of the normative analyzer, and that requires inverting, I believe):

echo 'Composing spelling relaxation transducer with normative analyser transducer to create descriptive analyser.'

hfst-compose -F -1 crk-orth.hfst -2 crk-anl-norm.hfst | hfst-minimize - -o crk-anl-desc.hfst

If you create it using the standard Giella Make file, I believe that knows how to properly compile (and invert when necessary).

The one annoying difference between HFST vs. XFST and FOMA is in how they interpret input/surface/lower and output/underlying/upper sides in some compositions (which I've never fully understood - to me, lookup means looking up, so the input would be the lower side and the analysis the upper side, which is at the same time the underlying form, which gets confusing). From folks in the know, the HFST solution would be more proper, but that's less of a consolation when XFST and FOMA operate differently.

from plains-cree-fsts.

We will probably want to implement a spell-relax rule also for word-final -uh, so that one can recognize neeyuh for nîya, and maybe some other interference from English.

from plains-cree-fsts.

The latest spellrelax file and how it is compiled (from crk.hfscript) should be the following:

hfst-regexp2fst -S -i $GTLANG_crk/src/orthography/spellrelax.regex | hfst-invert -o crk-orth.hfst

As far as I can remember, I've tried to have only one spellrelax file for all purposes.

For the record, for the spell-checker, I've also tried to specify exactly the same modifications for the error model, though that requires using the different formalism in the *.default.txt files.

from plains-cree-fsts.

It needs to be INVERTED?

from plains-cree-fsts.

These are the results I get when the spell relax is INVERTED:

echo "acuhkos" | flookup crk-descriptive-analyzer.fomabin

acuhkos	acâhkos+N+A+Sg
acuhkos	ahcahk+N+A+Der/Dim+N+A+Sg
acuhkos	atâhk+N+A+Der/Dim+N+A+Sg

echo "neeya" | flookup crk-descriptive-analyzer.fomabin

neeya	niya+Pron+Pers+1Sg
neeya	niyâ+Ipc

echo "neeyu" | flookup crk-descriptive-analyzer.fomabin

neeyu	niyâ+Ipc
neeyu	niya+Pron+Pers+1Sg

Does this look right?

from plains-cree-fsts.

Yup. I'm relying on the standard descriptive HFST in giella (our own script doesn't incorporate the regex's concerning various permutations of capitalization, but that shouldn't matter here):

hfst-optimized-lookup -q src/analyser-gt-desc.hfstol

You should also get the same results with ucukos and ucuhkos.

from plains-cree-fsts.

I inverted the FST as of 751f86f and now it works! Closing.

from plains-cree-fsts.