Phrase Based Models and Example Based Translation
A precursor to statistical machine translation, example based translation also attempts to construct translation by re-using parts of of translations in a pre-existing parallel corpus. The main points of difference is a less developed (or non-existant) probabilistic model and the goal to re-use as large chunks as possible.
Phrase Based Vs EBMT is the main subject of 12 publications. 8 are discussed here.
Phrase-based SMT is related to example-based machine translation (Somers, 1999)
. Some recent systems blur the distinction between the two fields (Groves and Way, 2005
; Paul et al., 2005
; Tinsley et al., 2008)
. Various combinations of methods from SMT and EBMT are explored by Groves and Way (2006)
. Statistical machine translation models may be used to select the best translation from several example-based systems (Paul and Sumita, 2006)
. Along these lines, phrase-based models may be improved with dynamically constructing translations for unknown phrases by using similar phrases that differ in a word or two and inserting lexical translations for the mismatched words (He et al., 2008)
Similar convergence takes place when combining statistical machine translation with translation memory, for instance by looking for similar sentences in the training data and replacing the mismatch with translation chosen with statistical translation methods (Hewavitharana et al., 2005)
- Murakami et al. (2010)
- Ma et al. (2011)
- Murakami et al. (2009)
- Wu (2005)