Apertium-uzb-kaa

Work plan

Week	Dates	Stems goal	Achieved?	Stems	Cov uzb	Cov kaa	Corpus testvoc uzb→kaa (no *, no */@, no */@/# errors)	Corpus testvoc kaa→uzb (no *, no */@, no */@/# errors)	WER,PER uzb→kaa	WER,PER kaa→uzb	Evaluation	Notes
0	20/05—26/05	--	--	(kaa) 5757 (uzb) 4083 (uzb-kaa) 412	jam(75.22) udhr(34.21) bible(78.80)	jam(95.37) bible(68.02)	jam(17.7, 17.7, 14.8) udhr(6.1, 6.1, 4.1) bible(9.6, 9.6, 8.2)	jam(34.1, 33.5, 16.2) udhr(n.a. in kaa) bible(17.0, 16.7, 9.7)	jam(96.19, 93.84 TODO properly) bible(TODO)	jam(106.15, 102.91 TODO properly) bible(TODO)
1	27/05—02/06	5000 from rus-{uzb,kaa} to each monodix	✘	(kaa) 7115 (uzb) 5934 (uzb-kaa) 412	jam(79.61) udhr(34.21) bible(79.58)	jam(95.37) bible(68.18)						uzb: 6c3887d kaa: fbb4425
2	03/06—09/06	5000 from rus-{uzb,kaa} to each monodix	✓
3	10/06—16/06	5000 from rus-{uzb,kaa} to each monodix	✓
4	17/06—23/06	5000 from rus-{uzb,kaa} to each monodix	✓
5	24/06—30/06	Corpus scraping		(kaa) 25513 (uzb) 34358 (uzb-kaa) 412	jam(85.31) udhr(39.47) bible(92.63)	jam(96.57) bible(93.01)						uzb: 432af5d kaa: 7a02618 uzb-kaa: c5ea3c3
13	19/08-25/08	Bidix		(kaa) 25512 (uzb) 34370 (uzb-kaa) 19363	jam(81.80) udhr(39.47) bible(94.64)	jam(96.57) bible(93.18)	jam(49.7, 49.7, 34.1) udhr(62.0, 60.1, 36.0) bible(49.1, 48.4, 30.3)	jam(72.9, 71.4, 45.2) udhr(n.a. in kaa) bible(56.0, 53.9, 38.8)	jam(69.74, 65.99) bible(112.11, 97.13)	jam(87.58, 84.16) bible(121.10, 104.57)		uzb: 4c121f4 kaa: 02f120d uzb-kaa: 25d79f6

Comments

• Bidix: number of stems in the bilingual dictionary.
  • Number of stems is taken from the header "aperium-dixtools format-1line 10 55 apertium-uzb-kaa.uzb-kaa.dix" produces.
• Corpus testvoc: cat <txt> | apertium-uzb-kaa/testvoc/corpus/trimmed-coverage.sh {uzb-kaa,kaa-uzb}. Corpora can be found here.
• WER is calculated like shown below. Currently WER numbers are only a crude approximation, since input text and -ref text are probably independent translations (not parallel).
• Evaluation is taking $n$ words and performing an evaluation for post-edition word error rate (WER). The output for those $n$ words should be clean.

Related terms:

• Trimmed coverage means the coverage of both of the morphological analysers after being trimmed according to the bilingual dictionary of the pair, that is, only containing stems which are also in the bilingual dictionary.
• Testvoc for a category means that the category is testvoc clean, in both translation directions.

apertium-uzb-kaa$ perl ../../../../../../src/sourceforge-apertium/trunk/apertium-eval-translator/apertium-eval-translator.pl -test <(cat ../../../data4apertium/corpora/jam/uzb.txt | apertium -d . uzb-kaa) -ref ../../../data4apertium/corpora/jam/kaa.txt
Test file: '/dev/fd/63'
Reference file '../../../data4apertium/corpora/jam/kaa.txt'

Statistics about input files
-------------------------------------------------------
Number of words in reference: 341
Number of words in test: 309
Number of unknown words (marked with a star) in test: 284
Percentage of unknown words: 91.91 %

Results when removing unknown-word marks (stars)
-------------------------------------------------------
Edit distance: 325
Word error rate (WER): 95.31 %
Number of position-independent correct words: 26
Position-independent word error rate (PER): 92.38 %

Results when unknown-word marks (stars) are not removed
-------------------------------------------------------
Edit distance: 328
Word Error Rate (WER): 96.19 %
Number of position-independent correct words: 21
Position-independent word error rate (PER): 93.84 %

Statistics about the translation of unknown words
-------------------------------------------------------
Number of unknown words which were free rides: 3
Percentage of unknown words that were free rides: 1.06 %