Samest meeting 27.9.2016

Participants: Fran, Heiki, Heli, Jaak, Jack, Sjur, Sulev, Tiina, Trond


  • New Oahpa
  • MT and CG
  • Estonian FST
  • Estonian Oahpa
  • Võro FST and Oahpa
  • Information
  • Next meeting

New Oahpa

Goals: modularisation, new design, path for learners, also with specific topics

apertium-apy: http://wiki.apertium.org/wiki/Apertium-apy

Comments: Lookupserver does the same (keeps process idle with no need for starting up process for each query) for fsts, but this apertium-apy promises doing the same for cg as well. Keep this in mind.

(also does load balancing and stuff)

Estonian FST

Heiki has worked on his FST version (experimental-langs), but not finished yet.

Normative/Descriptive fsts

-norm- and -desc- will differ when +Use/Err etc.

echo president+N+Sg+Ill|destNorm president+N+Sg+Ill presidendisse president+N+Sg+Ill presidenti

echo president+N+Sg+Ill|hfst-lookup tools/mt/apertium/generator-mt-gt-norm.hfstol > president+N+Sg+Ill presidenti 0.000000

The Russian FST has lots of Err/Orth, relegating forms that according to Zaliznyak are marginal.

Dialectal fsts

Generating the "dialect" forms:

file configure.ac:

# Specify the tags for all dialects in this variable, leave it empty if you do NOT support dialectal variant fst's. Use upper case, separate with space.
# Dialects are presently only used in Oahpa fst's, and only support dialectal variation within the -norm- fst's.
AC_SUBST([DIALECTS], ["Ord Rare"])

Documentation (for North Sámi)

  • Standard (Ordinary) Estonian = +Dial/+Ord, +Dial/-Rare
  • Traditional (Rare) Estonian = +Dial/+Rare, +Dial/-Ord
  • Unmarked strings will turn up in both versions

The tags above must be declared in root.lexc and the lexc must correspond to the tags specified in configure.ac (as said above). The tag may occur anywhere in the string, strings containing them are deleted.

There is nothing to do in any Makefile.am.

Estonian Oahpa

No progress recently.

Võro FST and Oahpa

Sulev has completed correction of the Jack's nouns' table (devtools). Now working with the verb table.

Gave sounds of Eesti-Võro to Heli. Heli will add them to the Vro-Oahpa vocabulary to be used in Leksa.

Heli is trying to add Võro synthetic voices to the Morfa C sentenses.

Vro synthetic voice: http://wi.ee/voro-kiil/voro-keele-sunteeshelu/ More technical information can be asked from Indrek Kiissel (ikiissel@gmail.com).

Jack has brought the recognition level for the Võro fst up to 32% testing vro wikitxt. Intending to work in Tartu with Sulev 4th through 6th of October, leaving 7th. (Oahpa-fst development and poster)

466 out of 2548 vro entries in vro-oahpa are not recognised by the analyser.

MT and CG

Tiina has been harmonizing conversion of tags to apertium format for pairs fin-est and fin-sme, this also enabled to simplify some transfer rules. Have some trouble with general conversion pipeline to apertium format:

2 of the tag conversion files from langs/lll/tools/mt/apertium/tagsets are compiled with the base transducer in different order for analysers compared to generators. The generators are composed with modify-tags.regex and then relabelled with apertium.postproc.relabel (as says doc), but the analysers are first relabelled and then composed with modify-tags.hfst. This may result in different tag usage in generators and analysers if to be unaware of such order difference.

Tiina will add this bug to Bugzilla, see http://giellatekno.uit.no/bugzilla.

Two issues:

  1. Different (non-corresponding) forms, hence different tags
  2. Different tag string for corresponding forms

Different (non-corresponding) forms, hence different tags

  • This should preferably be fixed in the apertium-lang1-lang2.lang1-lang2.t?x file.
  • Alternatively, we might e.g. generate +V+Neg+Sg1:ei etc for Sg2, for Estonian, etc.
  • In cg, we might disambiguate e.g. sme +Loc as @loc-ine vs. @loc-ela for subsequent translation to fin, sma or smj.


fin2est en, et: .t1x If you need context: do it in cg

This could be one-to-many, many-to-one, or just mismatch

Different tag string for corresponding forms

This should preferably be fixed in modify-tags.regex.

Examples of tag issues:

est: nad        nemad<prn><pl><nom>        0.000000 ***
est: tulema        tulema<vblex><actv><sup><ill>        0.000000
???: tulema        tulema<vblex><actv><infma>        0.000000
est: tulla        tulema<vblex><inf>        0.000000
est: ole        olema<vblex><actv><pres><imp><p2><sg>        0.000000
est: mina        mina<prn><sg><nom>        0.000000 ***
est: pole        pole+?        inf


fin: he        he<prn><pers><p3><pl><nom>        0.000000
fin: ne        ne<prn><dem><pl><nom>        0.000000
fin: ne        ne<prn><pl><nom>        0.000000
fin: tulla        tulla<vblex><actv><infa><sg><lat>        0.000000
???: tulla        tulla<vblex><actv><infa>        0.000000
fin: ole        olla<vblex><actv><indic><pres><conneg>        0.000000
fin: minä        minä<prn><pers><p1><sg><nom>        0.000000


fkv: he        he<prn><pers><p3><pl><nom>        0.000000
fkv: tulla        tulla<vblex><actv><infa><sg><lat>        0.000000
fkv: ole        olla<vblex><conneg>        0.000000


sme: boahtit        boahtit<vblex><iv><inf>        0.000000
sme: sii        son<prn><pers><p3><pl><nom>        0.000000
sme: dat        dat<prn><dem><pl><nom>        0.000000
sme: dat        dat<prn><dem><sg><nom>        0.000000
sme: leat        leat<vblex><iv><indic><pres><conneg>        0.000000

Should we add versions http://gtweb.uit.no/mt/testing/? No, not until hashtag-errors under 1%. hashtag errors are generation errors.

  1. 1<num><card><digit> KUS JAAKKO ON?
  2. 2<num><card><digit> Jaakko ja Mari on aias. Ilm on täna hea, on väga sooja. Aga eile oli väga külma! Siis #nemad<prn><pers><p3><pl><nom> ei #saama<vblex><actv><pret><indic><conneg> mängida väljas. Jaakko ja Mari peavad väga mängimisest, #nemad<prn><pers><p3><pl><nom> mängivad alati ühes aias suure maja ees.
  3. 3<num><card><digit> Jaakko on väike poiss ja #tema<prn><pers><p3><sg><nom> on kuus aastat vana. Väike tüdruk on #tema<prn><pers><p3><sg><gen> #nemad<prn><pers><p3><pl><gen> õe, #tema<prn><pers><p3><sg><nom> on viis aastat vana. Jaakkol on väike koer, ka koer on nüüd aias. Koerast on meeldivat mängida nende kahe lapsega. Koer on väga õnnelik nüüd.
  4. 4<num><card><digit> kas On ka Maril koer? Ei, Maril ei #olema<vblex><actv><pres><indic><conneg> koera, #tema<prn><pers><p3><sg><ade> on kass. Aga kass on majas, kass on magamas.

MT meeting this week: Tiina, Trond, Fran, Heiki-Jaan


  • Käbi Suvi and Keit Mõisavald are not actively working on MT at the moment. Status: sme-est.dix
  • There is some interest in Literary Museum (transcription or translation of Setu folk tales into Estonian).

Next meetings

  • MT-meeting: Friday 30. September, 10:00 Norwegian time
  • Samest-meeting: Tuesday, 18. October at 10:00 Norwegian time