Search Descriptions


Neural machine Translation

Statistical Machine Translation

Search Publications





Syntactic Pre-Reordering

Given the limitations of the dominating phrase-based statistical especially with long-distance reordering for syntactic reasons, it may be better to treat reordering in pre-processing.

Syntactic Prereordering is the main subject of 72 publications. 27 are discussed here.


Syntactic pre-reordering as a hand-crafted component has been explored for German–English (Collins et al., 2005), Japanese–English (Komachi et al., 2006). Wang et al. (2007) develop rules based on a constituent parser for Chinese–English, and then Cai et al. (2014) achieve slightly better performance with rules based on dependency parses. Ramanathan et al. (2008) reorder the verb for English–Hindi. Xu et al. (2009) define rules for the translation from English into several SOV languages (Korean, Japanese, Hindi, Urdu, Russian, Turkish). Carpuat et al. (2010); Carpuat et al. (2012) use a dependency parser to detect verb-subject constructions in Arabic and reorder them for translation into English.
Zwarts and Dras (2007) point out that translation improvements are due to both a reduction of reordering needed during decoding and the increased learning of phrases of syntactic dependents. Nguyen and Shimazu (2006) also use manual rules for syntactic transformation in a preprocessing step.
Such a reordering component may also be learned automatically from parsed training data, as shown for French–English (Xia and McCord, 2004). For Arabic–English, Habash (2007) learns reordering rules from a dependency parse and applies the most likely reordering pattern deterministically. Crego and Mariño (2007) learn rules for Chinese–English. Li et al. (2007) propose a maximum entropy pre-reordering model based on syntactic parse trees in the source language. Dyer and Resnik (2010) learn permutations for context free grammar rules and pass a reordered forest to a phrase-based system.
Genzel (2010) learns a sequence of reordering rules based on the dependency parse. The rules are learned greedily to reduce IBM Model 1 word alignment crossings. Jehl et al. (2014) learn pairwise child swapping preferences in the dependency tree. Gispert et al. (2015) use a feed-forward neural network for the same model.
Barone and Attardi (2013) develop a model of walking through the dependency structure to read out the source sentence in target side order and train a classifier to make decisions at each node, but fail to show improvements. Using a recurrent neural network to model this process (Barone and Attardi, 2015) is more successful.
Multiple reorderings may be encoded in a input lattice to the decoder make up for errors due to parse difficulty. This is shown to help for Chinese–English (Crego and Mariño, 2007). Bisazza and Federico (2010) focus on verb reordering in Arabic-English using a syntactic chunker, by moving the verb over varying number of syntactic chunks to the right. Andreas et al. (2011) use a dependency parser for the same purpose. To avoid the added decoding complexity of an input lattice, Bisazza and Federico (2012) encode reordering distances in a distortion matrix.
It may be beneficial to train different such pre-reordering models for different sentence types (questions etc.) (Zhang et al., 2008).
Preprocessing the input to a machine translation system may also include splitting it up into smaller sentences (Lee et al., 2008).


Khapra et al. (2012); Khapra et al. (2012b) organized a shared task on the pre-reordering problem.


Related Topics

New Publications

  • Na and Lee (2013)
  • Plamada et al. (2015)
  • Daiber et al. (2016)
  • Fuji et al. (2016)
  • Sudoh and Nagata (2016)
  • Takeno et al. (2016)
  • Oda et al. (2016)
  • Nakagawa (2015)
  • Hoshino et al. (2015)
  • Barone and Attardi (2015)
  • Herrmann et al. (2013)
  • Han et al. (2013)
  • Patel et al. (2013)
  • Hoshino et al. (2013)
  • Bisazza and Federico (2013)
  • Hatakoshi et al. (2014)
  • Stymne et al. (2014)
  • Ding et al. (2014)
  • Wu et al. (2014)
  • Lerner and Petrov (2013)
  • Gispert et al. (2014)
  • Bisazza and Federico (2013)
  • Dlougach and Galinskaya (2012)
  • Feng and Ney (2012)
  • Gupta et al. (2012)
  • Kunchukuttan and Bhattacharyya (2012)
  • Hu et al. (2014)
  • Khalilov and Sima'an (2012)
  • Jiang et al. (2010)
  • Navratil et al. (2012)
  • Lee et al. (2010)
  • Sudoh et al. (2010)
  • Arora and Sinha (2012)
  • Han et al. (2012)
  • DeNero and Uszkoreit (2011)
  • Nguyen and Shimazu (2006)
  • Holmqvist et al. (2009)
  • Li et al. (2009)
  • Elming and Habash (2009)
  • Khalilov and Sima'an (2010)
  • Isozaki et al. (2010)
  • Badr et al. (2009)
  • Jiang et al. (2010)
  • Katz-Brown et al. (2011)
  • Howlett and Dras (2011)