Search Descriptions

General

Neural machine Translation

Statistical Machine Translation

Search Publications


author

title

other

year

Parameter Tuning

Typical statistical machine translation systems are made up of several models which are weighted according to their importance. The estimation of these weights is called parameter tuning.

Parameter Tuning is the main subject of 47 publications. 10 are discussed here.

Topics in MachineLearning

Word Graphs | Parameter Tuning | Reranking | Large Scale Discriminative Training | Minimum Bayes Risk | Confidence Measures | System Combination | Incremental Updating

Publications

The popular minimum error rate training (MERT) method was proposed by Och (2003). Properties of this method are discussed by Moore and Quirk (2008). Cer et al. (2008) discuss regularizing methods that average BLEU scores over a small window to smooth over the curves.
Other researchers report on using the Simplex algorithm (Nelder and Mead, 1965). The implementation of these methods is described by (Press et al., 1997). Tuning may use multiple reference translations, or even additional reference translations generated by paraphrasing the existing reference (Madnani et al., 2007).
Traditionally, optimization is done on n-best lists of multiple decoder runs. Macherey et al. (2008) introduce a method to carry out MERT on lattices generated from the search graph. Sokolov and Yvon (2011) propose a faster algorithm by framing the problem in terms of semiring operations and using finite state machine toolkits.
One problem with MERT is instability: multiple tuning runs may lead to different weights and also differences in quality when evaluating a test set. Clark et al. (2011) argue that to obtain reliable results, MERT should be run three times and the results from the run with median test scores should be reported. Cettolo et al. (2011) suggest to run MERT multiple times and average the resulting weights. They also propose a method to run system combination methods to combine the output systems obtained from multiple MERT runs.

Benchmarks

Discussion

Related Topics

New Publications

  • Lo and Wu (2013)
  • Kocur and Bojar (2016)
  • Lo et al. (2015)
  • Zhai et al. (2015)
  • Dakwale and Monz (2016)
  • Neubig and Watanabe (2016)
  • Dreyer and Dong (2015)
  • Guzmán et al. (2015)
  • Nakov et al. (2013)
  • Nakov et al. (2013)
  • Lo et al. (2013)
  • Cer et al. (2013)
  • Green et al. (2014)
  • Galley et al. (2013)
  • Liu and Huang (2014)
  • Chiang et al. (2008)
  • Eidelman et al. (2013)
  • He and Way (2010)
  • Liu et al. (2011)
  • Zheng et al. (2011)
  • Nakov et al. (2012)
  • Chen et al. (2012)
  • Liu et al. (2012)
  • Cherry and Foster (2012)
  • Gimpel and Smith (2012)
  • Watanabe (2012)
  • Madnani et al. (2008)
  • Zaidan and Callison-Burch (2009)
  • Zhao and Chen (2009)
  • Foster and Kuhn (2009)
  • Utiyama et al. (2009)
  • He and Way (2009)
  • Chatterjee and Cancedda (2010)
  • UNKNOWN CITATION 'iwslt04:TP_cettolo'
  • Cong et al. (2010)
  • Sanchis-Trilles and Casacuberta (2010)
  • Galley and Quirk (2011)
  • Hopkins and May (2011)

Actions

Download

Contribute