Kazakh and Tatar/Work plan
Jump to navigation
Jump to search
This was a workplan for development efforts for Kazakh and Tatar translator in Google Summer of Code 2012.
- 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.
- Evaluation is taking words and performing an evaluation for post-edition word error rate (WER). The output for those words should be clean.
("news" is tat.news.2005-2011_300K-sentences.txt.bz2, "wp" - kaz.fixed.wikipedia.2010_100k-sentences.txt.bz2 [1])
| Week | Dates | B-trimmed Cov. Goal | Reached (news, wp) | Testvoc | Completed | Evaluation | WER | Notes |
|---|---|---|---|---|---|---|---|---|
| 0 | 45% | ??, ?? | <postadv> <ij> |
500 words | 13.39 % | Preliminary evaluation. Translate the story total coverage and without diagnostics (in both directions). Get a baseline WER. Work on disambiguation, the morphological ambiguities in the story should be resolved. | ||
| 1 | 50% | ??, ?? | <num> <post> |
- | Basic numerals, and postpositions should be clean. | |||
| 2 | 53% | 56.7%, 57.4% | <cnjcoo> <cnjadv> <cnjsub> |
- | ||||
| 3 | 59% | 59.3%, 59.8% | <adv> <abbr> |
200 words | 29.51% | Work on disambiguation. | ||
| 4 | 63% | 59.3%, 59.9% | <prn> <det> |
No | - | |||
| 5 | 68% | 67.1%, 64.3%[2] | <adj> <adj><advl> |
- | Adjectives with attributive and adverbial function should be clean. Numerals + affixes should have their phonology solved. | |||
| 6 | 70% | 73.4%, 70.3% | <n> <num><subst> <np> <adj><subst> |
500 words | 23.79%[3] | Midterm evaluation. Work on disambiguation. | ||
| 7 | 73% | 73.5%, 70.4% | - | Worked on translation of kaz.crp.txt | ||||
| 8 | 75% | 76.2%, 72.7% | <prn> |
No | - | Pronouns should be testvoc clean. | ||
| 9 | 77% | 77.7%, 75.5% | - | 200 words | Work on disambiguation. | |||
| 10 | 80% | 79.1%, 76.8% | <v> |
No | - | Verbs should be testvoc clean. | ||
| 11 | 80% | 79.9%, 77.4% | - | - | Work on disambiguation. | |||
| 12 | 80% | all categories clean | 500 words | Final evaluation (see below). Tidying up, releasing |
Final evaluation
Sources of texts can be find here: README
Azattyq news
selimcan@achilles:~/src/apertium/incubator/apertium-kaz-tat$ perl ../../trunk/apertium-eval-translator/apertium-eval-translator.pl -test texts/4story.kaz-tat.02_09_2012.txt -ref texts/4story.kaz-tat.postedit.txt Test file: 'texts/4story.kaz-tat.02_09_2012.txt' Reference file 'texts/4story.kaz-tat.postedit.txt' Statistics about input files ------------------------------------------------------- Number of words in reference: 534 Number of words in test: 536 Number of unknown words (marked with a star) in test: 2 Percentage of unknown words: 0.37 % Results when removing unknown-word marks (stars) ------------------------------------------------------- Edit distance: 133 Word error rate (WER): 24.91 % Number of position-independent correct words: 413 Position-independent word error rate (PER): 23.03 % Results when unknown-word marks (stars) are not removed ------------------------------------------------------- Edit distance: 135 Word Error Rate (WER): 25.28 % Number of position-independent correct words: 411 Position-independent word error rate (PER): 23.41 % Statistics about the translation of unknown words ------------------------------------------------------- Number of unknown words which were free rides: 2 Percentage of unknown words that were free rides: 100.00 %
Wikipedia article
selimcan@achilles:~/src/apertium/incubator/apertium-kaz-tat$ perl ../../trunk/apertium-eval-translator/apertium-eval-translator.pl -test texts/5story.kaz-tat.01_09_2012.txt -ref texts/5story.kaz-tat.postedit.txt Test file: 'texts/5story.kaz-tat.01_09_2012.txt' Reference file 'texts/5story.kaz-tat.postedit.txt' Statistics about input files ------------------------------------------------------- Number of words in reference: 515 Number of words in test: 518 Number of unknown words (marked with a star) in test: Percentage of unknown words: 0.00 % Results when removing unknown-word marks (stars) ------------------------------------------------------- Edit distance: 61 Word error rate (WER): 11.84 % Number of position-independent correct words: 460 Position-independent word error rate (PER): 11.26 % Results when unknown-word marks (stars) are not removed ------------------------------------------------------- Edit distance: 61 Word Error Rate (WER): 11.84 % Number of position-independent correct words: 460 Position-independent word error rate (PER): 11.26 % Statistics about the translation of unknown words ------------------------------------------------------- Number of unknown words which were free rides: 0 Percentage of unknown words that were free rides: 0%
Notes
- ↑ Despite the fact that both lexicons have the same amount of stems, transducer for Tatar usually has a lower coverage. This is probably because some forms aren't implemented yet, e.g. some verbal forms. Here I give numbers for both directions. Note that domains are different.
- ↑ As of Tuesday 26th June
- ↑ apertium-kaz-tat$ perl ../../trunk/apertium-eval-translator/apertium-eval-translator.pl -t texts/3story.kaz-tat.12072012.txt -r texts/3story.kaz-tat.postedit.txt
