Japanese-English Patterns

We developed the following process for forming Japanese-English patterns.

1. Parallel Japanese-English Corpus

   We prepare Japanese-English parallel sentences for training. Example sentences are listed in Table 1.

   
   

   Table: Parallel Japanese-English Corpus

   Japanese sentence	信号 は 赤 だ 。
   English sentence	The light was red .

   
   

2. Japanese-English Phrase Table

   We construct a Japanese-English phrase table using train-model.perl [Koehn
   
   2007]. An example Japanese-English phrase table is given in Table 2.

   
   

   Table: Example of Japanese-English Phrase Table

   Ex.1	信号 |||The light |||0.5 0.07 0.5 0.2
   Ex.2	信号 は |||lights |||0.01 0.06 0.03 0.04
   Ex.3	赤 |||red |||0.2 0.1 0.2 0.3
   Ex.4	赤 だ|||red |||0.3 0.2 0.2 0.2

   
   

3. Japanese-English High Probability Phrase Table

   We deleted the low-probability Japanese-English phrase table (Table 2), in which the threshold was 0.1. We call the resulting table a Japanese-English high-probability phrase table (HPPT). An example of an HPPT is presented in Table 3.

   
   

   Table: Example of Japanese-English High Probability Phrase Table

   Ex.1	信号 |||The light |||0.5 0.07 0.5 0.2
   Ex.3	赤 |||red |||0.2 0.1 0.2 0.3
   Ex.3	赤 だ|||red |||0.3 0.2 0.2 0.2

   
   

4. Japanese-English Patterns

   We used Japanese-English parallel sentences (Table ) and the Japanese-English HPPT (Table 3) to form Japanese-English patterns. Note that all possible Japanese-English patterns were generated. Therefore, one or more Japanese-English patterns were generated from one Japanese-English parallel sentence. Example Japanese-English patterns are listed in Table .

   
   

   Table: Japanese-English Patterns

   Ex.1	Japanese pattern	X1 は X2 だ 。
   	English pattern	X1 was X2 .
   Ex.2	Japanese pattern	X1 は X2 。
   	English pattern	X1 was X2 .

   
   

Figure 1 shows a flowchart for forming Japanese-English patterns.

Figure: Formation of Japanese-English Patterns