Turkic MT Improvements GSoC2019 report

From Apertium
Revision as of 18:30, 26 August 2019 by Oğuz (talk | contribs)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigation Jump to search

The aim of this project was improving the following language pairs of Apertium: tur->uig, uzb->tur, kir->tur, tat->tur. The first phase was mostly improving the bidix and lexc items of each language pair. After that the focus was on tur-uig and uzb-tur. CG and transfer rules were written to improve the translation quality for these languages.

Commits[edit]

My commits on each pair can be found here.

Transfer[edit]

Transfer rules were written for tur->uig and uzb->tur, using Regression Tests. They can be found here: Uighur and Uzbek. These rules focused on what the machine could not already translate. To this end missing suffixation patterns, lexical items and disambiguation rules etc. were also added to relative dictionaries, along with the transfer rules to enable the translation.

Corpora and Coverage[edit]

L Wiki Bible
Tur-Uig 53505239 words, 82.3% cov 178233 words, 93.0% cov
Uzb-Tur 12730161 words, 80.8% cov 184447 words, 83.5% cov
Kir-Tur 11435418 words, 82.5% cov 184808 words, 92.0% cov
Tat-Tur 5792382 words, 86.4% cov 178220 words, 91.4% cov

Dictionaries[edit]

All dictionaries were improved in the first stage of the project, with the help of mentors on Kipchak languages. Most frequent unknown tokens from corpora of each language (mostly consisting of Wikipedia entries, Bible and Quran) were added. Around 800-1000 entries were added to each language pair's bidix.

Disambiguation[edit]

To correctly discern the lemma and the morphology so as to be translated correctly into the target language, Apertium uses Constraint Grammar (CG). As part of the project, CG rules were added where necessary. Uzbek and Turkish in particular received extensive attention in this regard. For a better translation same/similar CG rules were written for these pairs, if the rule didn't clash with the intrinsic patterns of a language.

WER[edit]

---Uzbek---

Test file: 'mattauzbtr.txt' Reference file 'mattaturk.txt'

Statistics about input files


Number of words in reference: 565 Number of words in test: 579 Number of unknown words (marked with a star) in test: 124 Percentage of unknown words: 21.42 %

Results when removing unknown-word marks (stars)


Edit distance: 177 Word error rate (WER): 31.33 % Number of position-independent correct words: 408 Position-independent word error rate (PER): 30.27 %

Results when unknown-word marks (stars) are not removed


Edit distance: 188 Word Error Rate (WER): 33.27 % Number of position-independent correct words: 397 Position-independent word error rate (PER): 32.21 %

Statistics about the translation of unknown words


Number of unknown words which were free rides: 11 Percentage of unknown words that were free rides: 8.87 %,


---Uighur---

Test file: 'matta1turuig.txt' Reference file 'matta1uygur.txt'

Statistics about input files


Number of words in reference: 565 Number of words in test: 572 Number of unknown words (marked with a star) in test: 22 Percentage of unknown words: 3.85 %

Results when removing unknown-word marks (stars)


Edit distance: 270 Word error rate (WER): 47.79 % Number of position-independent correct words: 308 Position-independent word error rate (PER): 46.73 %

Results when unknown-word marks (stars) are not removed


Edit distance: 270 Word Error Rate (WER): 47.79 % Number of position-independent correct words: 308 Position-independent word error rate (PER): 46.73 %

Statistics about the translation of unknown words


Number of unknown words which were free rides: 0 Percentage of unknown words that were free rides: 0.00 %


---Kirgiz---

Test file: 'mattakirtr.txt' Reference file 'mattaturkkir.txt'

Statistics about input files


Number of words in reference: 569 Number of words in test: 669 Number of unknown words (marked with a star) in test: 63 Percentage of unknown words: 9.42 %

Results when removing unknown-word marks (stars)


Edit distance: 389 Word error rate (WER): 68.37 % Number of position-independent correct words: 286 Position-independent word error rate (PER): 67.31 %

Results when unknown-word marks (stars) are not removed


Edit distance: 389 Word Error Rate (WER): 68.37 % Number of position-independent correct words: 286 Position-independent word error rate (PER): 67.31 %

Statistics about the translation of unknown words


Number of unknown words which were free rides: 0 Percentage of unknown words that were free rides: 0.00 %

---Tatar---

Test file: 'mattatattr.txt' Reference file 'mattaturktatar.txt'

Statistics about input files


Number of words in reference: 573 Number of words in test: 587 Number of unknown words (marked with a star) in test: 66 Percentage of unknown words: 11.24 %

Results when removing unknown-word marks (stars)


Edit distance: 218 Word error rate (WER): 38.05 % Number of position-independent correct words: 375 Position-independent word error rate (PER): 37.00 %

Results when unknown-word marks (stars) are not removed


Edit distance: 218 Word Error Rate (WER): 38.05 % Number of position-independent correct words: 375 Position-independent word error rate (PER): 37.00 %

Statistics about the translation of unknown words


Number of unknown words which were free rides: 0 Percentage of unknown words that were free rides: 0.00 %

Future Plans[edit]

Uzbek lexicon still needs to be improved. Analysis of Uzbek can be problematic because of the unusual alphabet of the language along with non-standard forms, which also needs further improvement. More lexical selection, disambiguation and transfer rules are needed to achieve a greater translation quality on all pairs.