Search Descriptions

General

Neural machine Translation

Statistical Machine Translation

Search Publications


author

title

other

year

System Combination

If multiple machine translation systems are available, we may want to combine their outputs, or even their internal representations in some way to combine their strengths.

System Combination is the main subject of 104 publications. 22 are discussed here.

Publications

System combination is also called Multi-Engine Machine Translation (Nomoto, 2003), and may take the form of a classifier that decides which system's output to trust (Zwarts and Dras, 2008). The prevailing idea is to combine the output of different systems without deeper integration of their architectures, which has been very successful in speech recognition (Fiscus, 1997) and follows a machine learning paradigm called ensemble learning. Bangalore et al. (2002) use a confusion network representation for the computation of consensus translations of different off-the-shelf systems, while Zaanen and Somers (2005) use a word graph and Jayaraman and Lavie (2005) use a decoder to search through possible combinations, possibly augmented with a phrase translation table (Eisele et al., 2008). Eisele (2005) compares a statistical and a heuristic method.
Rosti et al. (2007); Rosti et al. (2007b) propose the construction of confusion networks by aligning the different system outputs, which may be improved by an incremental alignment process (Rosti et al., 2008). Matusov et al. (2006) use IBM Models to align the words in the different system translations, while Karakos et al. (2008) suggest to limit alignments with the ITG constraint. Word alignment may be aided by using information about synonyms, which may be provided by Wordnet synsets (Ayan et al., 2008). Huang and Papineni (2007) use some of the internals of the translation systems, such as phrase segmentation, to guide combination. Macherey and Och (2007) look at various properties of system combination methods, such as the relative quality of the system and how well they correlate, and its effect on success. System combination may be also used to combine systems that use the same methodology, but are trained on corpora with different morphological preprocessing (Lee, 2006), or different word alignment methods (Kumar et al., 2007).
Based on work on consensus translation, DeNero et al. (2010) go beyond system translations and consider the search graph of multiple systems to carry out system combination.

Benchmarks

System combination was the task at the WMT evaluation campaigns (Callison-Burch et al., 2009; Callison-Burch et al., 2010; Callison-Burch et al., 2011), which provide standard test sets.

Discussion

Related Topics

New Publications

  • Marie and Fujita (2018)
  • Marie et al. (2018)
  • Razmara and Sarkar (2013)
  • Federmann et al. (2012)
  • Formiga et al. (2013)
  • Nicolai et al. (2015)
  • Rikters (2015)
  • Freitag et al. (2015)
  • Rubino et al. (2016)
  • Ma and McKeown (2015)
  • Razmara and Sarkar (2013)
  • Ghannay and Barrault (2014)
  • Xia et al. (2013)
  • Singla et al. (2014)
  • Singla et al. (2014)
  • González-Rubio and Casacuberta (2013)
  • González-Rubio and Casacuberta (2013)
  • Bechara et al. (2011)
  • Liu et al. (2011)
  • Ma and Mckeown (2011)
  • Suzuki (2011)
  • Du and Way (2010)
  • Heafield and Lavie (2010)
  • Béchara et al. (2012)
  • Parton et al. (2012)
  • Tufiş and Dumitrescu (2012)
  • Liu et al. (2012)
  • Razmara et al. (2012)
  • Devlin and Matsoukas (2012)
  • Alegria et al. (2008)
  • Schwartz (2008)
  • Federmann (2012)
  • Li and Wang (2012)
  • Okita et al. (2012)
  • Rosti et al. (2012)
  • Soricut and Narsale (2012)
  • Federmann (2012)
  • Ma and McKeown (2012)
  • Potet et al. (2012)
  • He et al. (2008)
  • Hildebrand and Vogel (2008)
  • Hildebrand and Vogel (2009)
  • Duan et al. (2009)
  • Feng et al. (2009)
  • He and Toutanova (2009)
  • Gispert et al. (2009)
  • Zhao and He (2009)
  • Chen et al. (2009)
  • Federmann et al. (2009)
  • Heafield et al. (2009)
  • Leusch et al. (2009)
  • Rosti et al. (2009)
  • Chen et al. (2009)
  • Li et al. (2009)
  • Liu et al. (2009)
  • Aikawa and Ruopp (2009)
  • Du et al. (2009)
  • Duan et al. (2010)
  • Xiao et al. (2010)
  • Barrault (2010)
  • Bicici and Kozat (2010)
  • Bicici and Yuret (2010)
  • Chen et al. (2010)