Introduction

Machine translation (MT) systems have been studied for a long time, and there are now three generations of this technology. The first generation was a rule-based translation method, and the second generation was an example-based machine translation method. Recently, the statistical machine translation method has become very popular. This method is based on statistics. Many versions of statistical machine translation models are available. An early model of statistical machine translation was based on IBM1 $\sim$ 5[1]. Recent statistical machine translation systems usually use phrase-based models.

However some problems arise with phrase-based statistical machine translation. One problem is as follows. Normally, a translation model requires a large parallel corpus. However, if we use a smaller parallel corpus, it results in many unknown words in the output translation. The second problem is that normally, an $N$ -gram model is used as a language model. However, this model consists of local language information and does not have grammatical information.

To solve these problems, we have developed a two-stage machine translation system. The first stage is a rule-based machine translation system, and the second stage is a normal statistical machine translation system. This idea was based on paper[3],[4],[5].

In Chinese-English translation, the first stage consists of Chinese-English rule-based machine translation. In this stage, we obtained "ENGLISH" sentences from Chinese sentences. We aim to achieve "ENGLISH" sentences that contain few unknown words and that are generally grammatically correct. However, these "ENGLISH" sentences have low levels of fluency and naturalness because they were obtained using rule-based machine translation. In the second stage, we used a normal statistical machine translation system. This stage involves "ENGLISH" to English machine translation. With this stage, we aim to revise the outputs of the first stage improve the naturalness and fluency.

We used SYSTRAN V6 for the first stage. We used general statistical machine translation tools for the second stage, such as "Giza++"GIZA++, "moses" [7], and "training-phrase-model.perl" [10]. We used these data and these tools and participated in the BTEC-CE, Challenge-CE, and Challenge-EC contests at IWSLT2009.

As a result of experiments, the proposed method was not so effective for these tasks. The BLEU score was not as good compared to the standard moses. However, the score was not optimized, and our method was still very promising. Thus, we will continue to develop the method and try again in the future.