next up previous
Next: Size of trainig parallel Up: Discussion Previous: Removal of long parallel

Unknown Words

Some words were not translated and generated as unknown words. Almost of all these words were a person's name or a place-name. Thus, if we add a procedure for unknown word, we will obtain better better. By the way, to get higher BLEU scores, a well-known trick in the field is to simply delete unknown words form the output. Table12 shows these results. To compare table11, the blue score improved about 1 persent for most experiment conditions.



Jin'ichi Murakami 2008-10-28