n-best Translation
We generally used 1-best translation
- Highest unigram frequency in comparable corpus
Tried 2-best: two highest-ranked translations
- Duplicating unique translations where necessary
-
Should reduce miss rate
- But at what cost in false alarms?