Decoding Pattern

The decoding process is as follows.

1. Input Japanese Sentences

   We prepare input Japanese sentences. An example sentence is given in Table 5.

   
   

   Table: Japanese Sentence

   郵便 局 は どこ に 有り ます か ？

   
   

2. Search Japanese Pattern and Output English Pattern

   We search for a Japanese pattern that is matched with the input Japanese sentence using Japanese patterns and the HPPT (section 4.1). Then we obtain English patterns. Example Japanese-English patterns are listed in Table 6. Also, an example Japanese-English HPPT is shown in Table 7.

   
   

   Table: Japanese-English Patterns

   Ex.1	Japanese	X1 X2 は どこ に 有り ます か ？
   	Pattern
   2-3	English	Where's the X2 X1 ?
   	Pattern
   Ex.2	Japanese	X2 は どこ に 有り ます か ？
   	Pattern
   2-3	English	Where is a X2 ?
   	Pattern

   
   

   
   

   Table: Japanese-English High-Probability Phrase Table

   Ex.1	局 |||post |||0.5 0.07 0.5 0.21
   Ex.2	郵便 |||postal |||0.4 0.031 0.2 0.11
   Ex.3	郵便 局 |||postal service |||0.1 0.07 0.1 0.01

   
   

3. Generate English Sentences

   We generate English sentences using the English pattern and Japanese-English High-Probability phrase tables. Note that all possible English sentences are generated. Therefore, multiple English sentences are generated from an input Japanese sentence. Example English sentences are listed in Table .

   
   

   Table: Generated English Sentences

   Ex.1	Where's the post office ?
   Ex.2	Where is a post station ?

   
   

4. Select English Sentence.

   We select one English sentence from the multiple generated English sentences using $3$ -gram. We used the n-gram-count in the Stanford Research Institute Language Model (SRILM) toolkit [Stolcke2002] and used ``-ukndiscount -interpolate'' as the smoothing parameter.

   An example of an English sentence that might be selected is shown in Table 9.

   
   

   Table: Select English Sentence

   Where is a post station ?

   
   

Figure: Decoding of Pattern-Based Machine Translation

Figure 2 shows the process of decoding English sentences for PBMT. We refer to the output of the proposed pattern-based machine translation as ``English'' sentences.