... Mori1, Tatsuya Kawahara11Graduate School of Informatics, Kyoto UniversityYoshida Honmachi, Sakyo-ku, Kyoto, Japan2National Institute of Information and Communication Technology3-5 Hikari-dai, ... note that only phrasesgenerated by Ptare actually memorized and givenhigher probability by the model. In FLAT, only min-imal phrases generated after Pxoutputs the terminalsymbol TERM are ... Kneser-Ney 5-gram model. For GIZA ++, we use the standard training reg-imen up to Model 4, and combine alignmentswith grow-diag-final -and. For the proposedmodels, we train for 100 iterations, and...