Finite State Machines
Statistical machine translation models may be implemented using finite state machines, for which a large number of powerful toolkits are available, which provide their own generic decoding algorithms.
Finite State Machines is the main subject of 18 publications. 10 are discussed here.
Instead of devising a dedicated decoding algorithm for statistical machine translation, finite state tools may be used, both for word-based (Bangalore and Riccardi, 2000
; Bangalore and Riccardi, 2001
; Tsukada and Nagata, 2004
; Casacuberta and Vidal, 2004)
, alignment template (Kumar and Byrne, 2003)
and phrase-based models. The use of finite state toolkits also allows for the training of word-based and phrase-based models. The implementation by Deng and Byrne (2005)
is available as the MTTK toolkit (Deng and Byrne, 2006)
. Similarly, the IBM models may be implemented using graphical model toolkits (Filali and Bilmes, 2007)
. Pérez et al. (2007)
compare finite state implementation of word and phrase-based models.
Just as word-based and phrase-based models may be implemented with finite state toolkits, a general framework of tree transducers may subsume many of the proposed tree-based models (Graehl and Knight, 2004)
- Argueta and Chiang (2017)
- Iglesias et al. (2009)
- González and Casacuberta (2009)
- Malik et al. (2010)
- Iglesias et al. (2011)
- Alshawi et al. (2002)
- Beck (2011)
- Vogel and Ney (2000)