|
Czech-English Translation
Petrželka, Jiří ; Schmidt, Marek (oponent) ; Smrž, Pavel (vedoucí práce)
This Master's thesis describes the principles of statistical machine translation and demonstrates how to assemble the Moses statistical machine translation system. In the preparation step, a research on freely available bilingual Czech-English corpora is done. An empirical analysis of time requirements of multithreaded word alignment tools demonstrates that MGIZA++ can achieve a five-fold speed-up, while PGIZA++ can reach an eight-fold speed-up (compared to GIZA++).Three scenarios of morphological pre-processing of Czech training data are tested, using simple unfactored models. While pure lemmatization can aggravate the BLEU, more sophisticated approaches usually raise BLEU. The positive effect of morphological pre-processing diminishes as corpus size rises. The relation between other corpora characteristics (size, genre, extra data) and the resulting BLEU are empirically gauged. A final system is trained on the CzEng 0.9 corpus and evaluated on the testing set from WMT 2010 workshop.
|
|
Systém pro kontrolu slovníků
Solanský, Petr ; Kouřil, Jan (oponent) ; Smrž, Pavel (vedoucí práce)
Práce je zaměřena na implementaci informačního systému pro kontrolu a opravu překladových a výkladových elektronických slovníků ve formátu LMF. Systém nabízí sedm typů kontrol a jednu opravu hromadně měnící obsahy slovníků. V technické zprávě jsou popsány nejdůležitější použité technologie, konceptuální návrh systému i kontrol samotných, důležité implementační prvky a výsledky se statistikou tohoto informačního systému.
|
|
Czech-English Translation
Petrželka, Jiří ; Schmidt, Marek (oponent) ; Smrž, Pavel (vedoucí práce)
This Master's thesis describes the principles of statistical machine translation and demonstrates how to assemble the Moses statistical machine translation system. In the preparation step, a research on freely available bilingual Czech-English corpora is done. An empirical analysis of time requirements of multithreaded word alignment tools demonstrates that MGIZA++ can achieve a five-fold speed-up, while PGIZA++ can reach an eight-fold speed-up (compared to GIZA++).Three scenarios of morphological pre-processing of Czech training data are tested, using simple unfactored models. While pure lemmatization can aggravate the BLEU, more sophisticated approaches usually raise BLEU. The positive effect of morphological pre-processing diminishes as corpus size rises. The relation between other corpora characteristics (size, genre, extra data) and the resulting BLEU are empirically gauged. A final system is trained on the CzEng 0.9 corpus and evaluated on the testing set from WMT 2010 workshop.
|