Experiments on Indonesian-Japanese statistical machine translation

Abstract

Based on the characteristics of Indonesian and Japanese language, we did several experiments on the additional process to an Indonesian-Japanese statistical machine translation (SMT). We proposed several additional processes such as employing the POS tag information, adding the size of monolingual target corpus, using Indonesian stemmer in Indonesian to Japanese translation, eliminating Japanese particle in Japanese to Indonesian translation, and the elimination of NE tag. The experimental result showed that compared to the baseline of adding no process to the default SMT engine (here, we use Moses), the highest BLEU score was achieved by the elimination of Japanese particle in Japanese to Indonesian translation.