User:Firespeaker/GSoC2014/Workplan

From Apertium
< User:Firespeaker‎ | GSoC2014
Revision as of 16:49, 16 June 2014 by Firespeaker (talk | contribs) (→‎Schedule: week 4 eval)
Jump to navigation Jump to search

Primary goals

  • A production-ready release of kaz-kir
    • Translates kaz→kir and kir→kaz with consistently <10% WER
    • Trimmed coverage for kaz and kir ≥90%
  • A production-ready release of tur-kir
    • Translates tur→kir and kir→tur with consistently <20% WER
    • Trimmed coverage for tur and kir ≥85%
  • A stable release of uzb-tur
    • Translates tur→uzb and uzb→tur with consistently <25% WER
    • Trimmed coverage for tur and uzb ≥80%
  • While bidix size is not built into the goals, the trimmed coverage numbers can be seen as a more relevant proxy for the same basic idea.

Plan

Schedule

See GSoC 2014 Timeline for complete timeline.

week dates goals eval accomplishments notes
post-application period
22 March - 20 April
  • apertium-kir to 90% coverage (with kaz-like transducer)
  • apertium-tur to 90% coverage (with kaz-like transducer)
  • apertium-uzb to 90% coverage (with kaz-like transducer)
  • build arsenal of texts with post-edited translations:
    • four 200-word texts in each kaz, kir, tur, uzb
    • four 500-word texts in each kaz, kir, tur, uzb
  • Reworked apertium-tur verb morphology on paper
  • Much better disam in apertium-tur
  • Have a bunch of texts, not many post-edited
  • Need to rework transfer rules for apertium-tur
community bonding period
21 April - 19 May
  • tur-uzb bidix to 7000 stems
  • make some baseline CG for kir, uzb
  • one 200-word kaz-kir text to <10% WER
  • one 200-word kir-kaz text to <10% WER
  • one 200-word tur-kir text to <10% WER
  • one 200-word kir-tur text to <10% WER
  • one 200-word tur-uzb text to <10% WER
  • one 200-word uzb-tur text to <10% WER
  • Implemented reworking of apertium-tur verb morphology
  • Around, but lost time due to end of semester and conferences
1 19 - 24 May
  • one 200-word kir-tur text to <10% WER
  • one 200-word kir-kaz text to <10% WER
  • work on kir CG and lrx
  • Fixed some apertium-tur phonology to go with morphology rework
  • Mostly worked on LREC poster
2 25 - 31 May
  • one 200-word kaz-kir text to <10% WER
  • one 200-word tur-kir text to <10% WER
  • work on kaz CG and lrx
  • work on tur CG and lrx
  • At LREC, worked a lot on apertium-uig and apertium-kaz-uig
3 1 - 7 June
  • one 200-word tur-uzb text to <10% WER
  • one 200-word uzb-tur text to <10% WER
  • work on uzb CG and lrx
  • work on tur CG and lrx
  • Fixed a small handful of transfer issues in apertium-tur-kir
  • Got apertium-uig in a state for others to work with
  • Completed eval for last few weeks
4 8 - 14 June
  • one 500-word kir-tur text to <10% WER
  • one 500-word kir-kaz text to <10% WER
  • work on kir CG and lrx
  • start testvoc nouns for all pairs
  • Brought coverage of apertium-tur-kir on SETimes corpus up by 2%
  • Brought testvoc of apertium-tur-kir on SETimes corpus down from 10.39% to 0.22%
5 15 - 21 June
  • one 500-word kaz-kir text to <10% WER
  • one 500-word tur-kir text to <10% WER
  • work on kaz CG and lrx
  • work on tur CG and lrx
  • continue testvoc nouns for all pairs
6 22 - 28 June
  • one 500-word tur-uzb text to <10% WER
  • one 500-word uzb-tur text to <10% WER
  • work on tur CG and lrx
  • work on uzb CG and lrx
  • continue testvoc nouns for all pairs
midterm eval
29 June
  • kaz(-kir) trimmed coverage ≥90%
  • kir(-kaz) trimmed coverage ≥90%
  • tur(-kir) trimmed coverage ≥90%
  • kir(-tur) trimmed coverage ≥90%
  • tur(-uzb) trimmed coverage ≥80%
  • uzb(-tur) trimmed coverage ≥80%
7 29 June - 5 July
  • finish testvoc nouns for all pairs
8 6 - 12 July
  • testvoc adjs for all pairs
9 13 - 19 July
  • testvoc numerals for all pairs
10 20 - 26 July
  • testvoc v.iv for all pairs
11 27 July - 2 August
  • testvoc v.tv categories for all pairs
12 3 - 10 August
  • testvoc adverbs for all pairs
  • testvoc misc categories for all pairs
pencils-down week
final evaluation
11 August - 18 August
  • move pairs to trunk
  • document stuff better on the wiki
  • make the pairs live at turkic.apertium.org

Getting started

  • make scripts for:
    • getting raw numbers for Progress
    • doing regression tests (/learn how to use existing scripts)
    • guesser
  • get updated corpora for:
    • Uzbek
    • Turkish

Recurring

  • The end of every week:
  • Constantly:
    • Add good sentences to regression tests
    • Clean up lexc files
      • remove duplicate entries
      • alphabetise sections?
      • add glosses, etc.