Introduction

Machine translation (MT) systems have been extensively studied, and there are now three generations of this technology. The first generation consists of rule-based MT (RBMT) systems. A pattern-based MT (PBMT) system is a kind of RBMT system. The second generation consists of example-based machine translation systems, and the third generation consists of statistical machine translation (SMT) systems, which have become very popular. Many versions of SMT systems have been introduced. An early SMT system was based on word-based models (IBM 1 $\sim$ 5[Brown

1993]). Recent statistical MT systems have usually used phrase-based models.

However, some problems arise with phrase-based SMT. One problem is the language model. Generally, an $N$ -gram model is used as the language model. However, this kind of model includes only local language information and does not include grammatical information. To solve this problem, we developed a two-stage MT system. The first stage consists of an automatically created PBMT system, and the second stage consists of a standard SMT system.

For Japanese-English translation, the first stage consists of Japanese-English PBMT. In this stage, we obtain ``English'' sentences from Japanese sentences. Our aim is to produce grammatically correct ``English'' sentences. However, these ``English'' sentences sometimes have low levels of fluency because they were obtained using an automatically created PBMT. In the second stage, we use a standard SMT system. This stage involves ``English'' to English machine translation. With this stage, our aim is to revise the outputs of the first stage in order to improve fluency.

We developed a PBMT system for the first stage using ``train-model.perl''[Koehn

2007]. We also developed a standard SMT system for the second stage using general SMT tools such as ``Moses'' [Koehn

2007]. We used these data and tools to translate Japanese-English simple sentences.

We obtained a Bilingual Evaluation Understudy (BLEU) score[Papineni

2002] of 0.1821 with our proposed system. In contrast, we obtained a BLEU score of 0.2218 in the Japanese-English simple sentences using a standard SMT system (Moses). This means that the proposed system was not effective for automatic evaluation in the Japanese-English simple sentence task.

However, we conducted ABX tests[Clark1982] to compare the output of the standard SMT system (Moses) and the output of the proposed system for 100 sentences. The results indicated that 30 sentences of the proposed system were thought to be better than those of the standard SMT system, and 9 sentences of the standard SMT system were thought to be better than those of the proposed system. This means that our proposed system was effective in the Japanese-English simple sentence task for human evaluation.