next up previous
Next: Pattern-Based Machine Translation Up: Two stage Machine Translation Previous: Concept of Two-Stage Machine

Related Work

Two-stage MT systems have been proposed before [Xu and Seneff2008], [Ehara2007], [Dugast
\bgroupet al.\end{tex2html_bgroup}
2007
], [Simar
\bgroupet al.\end{tex2html_bgroup}
2007
]. L. Dugast, et al. [Dugast
\bgroupet al.\end{tex2html_bgroup}
2007
] and M. Simard, et al. [Simar
\bgroupet al.\end{tex2html_bgroup}
2007
] applied SYSTRAN and SMT for Japanese-English translation. Their concept was to use SMT as a post-process for SYSTRAN. The results of these studies indicated that these systems are more effective than using SYSTRAN or SMT alone. In M. Simard's research [Simar
\bgroupet al.\end{tex2html_bgroup}
2007
], the BLEU score was 0.2598 for SMT and 0.2880 for SYSTRAN + SMT in English-Japanese translation, and 0.2517 for SMT and 0.2679 for SYSTRAN + SMT in Japanese-English translation. Ehara[Ehara2007] reported on the same system for Japanese-English translation of a patent task. The BLEU score was 0.2821 for SMT and 0.2921 for RBMT + SMT. Ehara's RBMT system was a commercial Japanese-English system. For these systems, SMT was used in the post-process for RBMT, which means that SMT was used as a means of language adaptation. Also, these RBMT systems were created by hand, so they were expensive to build.



Jin'ichi Murakami 2012-11-06