Přibližná extrakce frázové tabulky z velkého paralelního korpusu

The aim of this work is to examine the applicability of an algorithm for approximate frequency counting to act as an on-the-fly filter in the process of phrase table extraction in Statistical Machine Translation systems. Its implementation allows for the bulk of extracted phrase pairs to be much reduced with no significant loss to the ultimate quality of the phrase-based translation model as measured by the state-of-the-art evaluation measure BLEU. The result of this implementation is a fully working program, called eppex, capable of acting as an alternative to the existing tools for phrase table creation and filtration that are part of the open-source SMT system Moses. A substantial part of this work is devoted to the benchmarking of both the runtime performance and the quality of produced phrase tables achieved by the program when confronted with parallel training data comprised of 2 billions of words. Powered by TCPDF (www.tcpdf.org