160628

Participants: Heiki, Heli, Jaak, Sjur, Tiina, Trond

• Estonian FST
• Fin2X MT
• Next meetings

Heiki-Jaan has worked on the experimental-langs directory, converted filolex to giella format, and found some irregular issues calling for treatment, and has made some changes in the 2-level rules.

Jaak has looked at experimental-langs, so that it compiles for xfst also.

The message is: Something is moving and changing. We have two approaches to Estonian fst now, and try out both.

The fin analyser

• fin2X, x = est, sme.
• We made sme2fin before christmas:
  • fin dis was allegedly so bad
  • fin fst overgenerated
  • fin fst contains multiple path issues.

The situation now is:

• fin fst overgenerates, but only to some degree (not much).
• fin dis:
  • Development corpus of 2000 sentences, 15500 words. Result: 99.7%
  • Test corpus fiwiki of 61000 words: ambiguity: 1.14 readings/word
• fin fst multiple tag paths: Trivial corrections

The positive conclusion:

• Things are not that bad
• We should remove banalities to face the real MT issues.

TODO (Tiina, Trond)

Finish the fin fst cleanup. Again, have a look at the CG output + tag unification. The next focus could then be on MT output.

Fin2est

Bachelor student Keit Mõisavald will attend the summer school in Alacante, her thesis will be on translating fin clitics (-ko, -han, etc) into Estonian.

Issues in sme2smX:

• sme: ipmirdat go dan (norw) / ipmirdatgo dan (finn)
• smj: norw only
• smn: finn only

sme2xmX has to deal with both "norw2finn" and "finn2norw"

The Saami MTs have worked quite a lot with clitics, and will be relevant (as will eventual work on fin2sme or sme2fin).

Issues for MT:

• Scope of positive/negative -kin / -kAAn
• Clitic or adverbial (Pekkakin vs. myös Pekka)

Current "ambiguity":

4
"<tietenkin>" Pcle Pcle Foc/kin
        "tietenkin" Pcle
        "tieten" Pcle Foc/kin  <== of course plain wrong

Do we want to lexicalise the common light words + clitics or not? The answer that meets the demand of the long tail is the latter. Go through the presentations of 30 years of Virsu seminars?

• MT: 30.6. 0900 Norwegian time
• Samest: August 4th at 14.00 EET, 13.00 CET.

Sjur will probably still be on vacation on August 4.