Difference between revisions of "User:Eden/GSoC progress"

From Apertium
Jump to navigation Jump to search
 
(16 intermediate revisions by 2 users not shown)
Line 1: Line 1:
== Community Bonding Period ==
* Find Swahili-Lingala resources
* Update Lingala lexc transducer to lexd
* New lexd transducer for Swahili
* Keep track of coverage for Lin and Swa transducers
* Get familiar with apertium-recursive
* Set up <code>swa-lin</code> pair using apertium-recursive
* Update GSOC progress page
* James and Marry story + Wikipedia article in Swahili and Lingala.

== Goals ==
* By first evaluation: have story about kids or similar text to WER/PER of around 20% (work with all stages of translation, focus on "lowest-hanging fruit" relevant to the text)
* By second evaluation: increase [trimmed] coverage to around 90% (work focused on lexicons, adding from frequency lists)
* By final evaluation: work to get clean testvoc (work focused on transfer, making sure everything is dealt with one way or other)

== Status table ==
== Status table ==


Line 4: Line 19:
|-
|-
!colspan="2"|Week
!colspan="2"|Week
!colspan="2"|Stems
!colspan="3"|Stems
!colspan="2"|naïve coverage
!colspan="3"|naïve coverage
!colspan="2"|WER,PER
!colspan="2"|WER,PER
!colspan="2"|Progress
!colspan="2"|Progress
Line 11: Line 26:
! №
! №
! dates
! dates
! swa
! lin
! lin
! lin-eng
! swa-lin
! swa
! lin
! lin
! lin-eng
! swa-lin
! swa→lin
! lin→eng
! lin→swa
! eng→lin
!Evaluation
!Evaluation
!Notes
!Notes
|-
|-
| 0 (community bonding)
| 0
| May 20 - May 26
| May 4 - May 31
| 727
| 86
| 139
| 1,444
| 61.95%
| 26
| 40.86%
| 86.79%,80.87%
| 75.27%,63.98%
|
|
|
|-
| 1
| May 27 - June 02
| 904
| 139
| 62.57%
| 40.86%
| 86.79%,80.87%
| 75.27%,63.98%
|
|
|
|
|-
| 2
| May 03 - June 09
| 1,154
| 1,416
| 63.17%
| 53.03%
| 87.02%,79.95%
| 74.46%,60.22%
|
|
|-
| 3
| June 10 - June 16
| 1,172
| 1,501
|
| 61.60%
| 91.57%,79.04%
| 75.85%,62.90%
|
|
| WER for 'lin-eng' went up because of an incomplete rule for verbs that creates unnecessary pronouns. Main work next week will be on rules to dramatically improve WER and PER.
|-
| 4
| June 17 - June 23
| 1,200
| 1,540
| 69.70%
| 62.70%
| 79.27%,64.24%
| 84.41%,72.58%
|
|
|
|-
| 5
| June 24 - June 30
| 1,200
| 1,556
| 70.21%
| 61.90%
| 77.68%,67.88%
| 85.48%,73.92%
|
|
|
|-
|-
| 6
| 1
| July 1 - July 7
| June 1 - June 7
|
| 86
| 1,444
| 26
|
|
|
|
Line 96: Line 63:
|
|
|-
|-
| 7
| 2
| July 8 - July 14
| May 8 - June 14
| 170
|1,236
|1,577
| 1,444
| 26
|69.35%
|60.47%
|60.59%,46.47%
|72.61%,58.68%
|
|
|Work was done on lexical selection and rules about determinants. Current lexical selection works well with the text currently in use, which is a more rigid and literary Lingala. Further tests will be run on texts from the Wikipedia corpus to generalize lexical rules.
|-
| 8
| July 15 - July 21
|1,280
|1,580
|72.81%
|68.62%
|52.62%,42.82%
|59.04%,46.28%
|
|
|
| WER went down in both directions by approximately 2% after I added accents, and missing ɔ́ ɔ ɛ́ ɛ. Next focus will be on negation and trying to find a bigger corpus(>1000 words).
|
|
|
| Number of stems in lin transducer comes from prev. estimates. Manually counted stems in swa transducer
|-
|-
| 9
| 3
| July 12 - July 28
| June 15 - June 21
| 303
| 1,444
| 26
|
|
|
|
Line 126: Line 87:
|
|
|
|
| work was mainly collecting and finding stems.
|-
| 4
| June 22 - June 28
| 6,667
| 1,716
| 1,436
|
|
| 76.5%
|
|
| 94.40%
| 107.95%
|
| several duplicates in the swa transducer.
|-
| 5
| June 29 - July 5
|-
| 6
| July 6 - July 12
|-
| 7
| July 13 - July 19
|-
| 8
| July 20 - July 26
|-
| 9
| July 27 - Aug 2
|-
| 10
| July 3 - Aug 9
|-
| 11
| Aug 10 - Aug 16
|-
| 12
| Aug 17 - Aug 23
|-
|}
|}

== Work ==
* June 8 - June 14
- verb, noun, adjective morphotatics in swa transducer
* June 15 - June 21
- add missing verb TAM(continuative, reciprocal,causative)(<br/>
- more subsections in 'Verb Morphotatics'<br/>
- add stems in swa transducer <br/>
- start writing transfer rules <br/>


== Notes ==
== Notes ==

Latest revision as of 15:06, 27 June 2020

Community Bonding Period[edit]

  • Find Swahili-Lingala resources
  • Update Lingala lexc transducer to lexd
  • New lexd transducer for Swahili
  • Keep track of coverage for Lin and Swa transducers
  • Get familiar with apertium-recursive
  • Set up swa-lin pair using apertium-recursive
  • Update GSOC progress page
  • James and Marry story + Wikipedia article in Swahili and Lingala.

Goals[edit]

  • By first evaluation: have story about kids or similar text to WER/PER of around 20% (work with all stages of translation, focus on "lowest-hanging fruit" relevant to the text)
  • By second evaluation: increase [trimmed] coverage to around 90% (work focused on lexicons, adding from frequency lists)
  • By final evaluation: work to get clean testvoc (work focused on transfer, making sure everything is dealt with one way or other)

Status table[edit]

Week Stems naïve coverage WER,PER Progress
dates swa lin swa-lin swa lin swa-lin swa→lin lin→swa Evaluation Notes
0 (community bonding) May 4 - May 31 86 1,444 26
1 June 1 - June 7 86 1,444 26
2 May 8 - June 14 170 1,444 26 Number of stems in lin transducer comes from prev. estimates. Manually counted stems in swa transducer
3 June 15 - June 21 303 1,444 26 work was mainly collecting and finding stems.
4 June 22 - June 28 6,667 1,716 1,436 76.5% 94.40% 107.95% several duplicates in the swa transducer.
5 June 29 - July 5
6 July 6 - July 12
7 July 13 - July 19
8 July 20 - July 26
9 July 27 - Aug 2
10 July 3 - Aug 9
11 Aug 10 - Aug 16
12 Aug 17 - Aug 23

Work[edit]

  • June 8 - June 14

- verb, noun, adjective morphotatics in swa transducer

  • June 15 - June 21

- add missing verb TAM(continuative, reciprocal,causative)(
- more subsections in 'Verb Morphotatics'
- add stems in swa transducer
- start writing transfer rules

Notes[edit]

  • To count stems in lexc, try:
 grep -E ":\w+.*;" apertium-lin.lin.lexc | grep -v "[<>]" | wc -l
  • To count stems in the bidix, try this:
 grep "<p" apertium-eng-lin.eng-lin.dix  | wc -l
  • To get WER and PER use apertium-eval-translator-line