We conducted additional experiments. We replaced PBMT with other machine translation systems. Table 12 summarizes the results of these experiments for the BTEC-FE task. The components of the tables are as follows.
"Proposed" means our proposed system.
"MOSES" means a standard SMT with parameter tuning.
"MOSES (no tuning)" shows a standard SMT with no parameter tuning.
"SYSTRAN + MOSES" means the first stage is SYSTRAN and the second stage is a standard SMT with parameter tuning.
"SYSTRAN + MOSES (no tuning)" means the first stage is SYSTRAN and the second stage is a standard SMT with no parameter tuning.
"PBMT + MOSES" means the proposed system. This means the first stage is automatically created PBMT and the second stage is a standard SMT with parameter tuning.
"PBMT + MOSES (no tuning)" means the first stage is automatically created PBMT and the second stage is a standard SMT with no parameter tuning.
"JOSHUA" means a standard hierarchical phrase-based machine translation (HPMT)[13].
"SYSTRAN + JOSHUA" means the first stage is SYSTRAN and the second stage is HPMT [13].
"SYSTRAN + MOSES" (first stage SYSTRAN and second stage SMT with parameter tuning) seems to be the best system for many scores (BLEU, METEOR, etc.). SYSTRAN uses additional language resources. Therefore, these outputs have less unknown words, and there are few ungrammatical sentences.