Machine Translation without Word Alignments
While the vast majority of statistical machine translation systems maintain an alignment of source and target words (or word groups), some work explored machine translation without such alignment.
MT Without Word Alignment is the main subject of 5 publications.
While word alignment is generally assumed to be an essential element of statistical translation models, methods have been proposed that first generate a bag of words from the source bag of words and then apply an ordering model (Venkatapathy and Bangalore, 2007
; Bangalore et al., 2007)
Alignment-free word translation models may be used as a feature function in traditional models. Mauser et al. (2009)
employs a maximum entropy model that predicts output words from all source words as an additional scoring function. Huck et al. (2011)
confirm the effectiveness of the method.