Search Descriptions

General

Neural machine Translation

Statistical Machine Translation

Search Publications


author

title

other

year

Confidence Measures

For various applications of machine translation, it is useful to know how good the translation of a given document, sentence or even word is. Recently, the concept of a system assessing its confidence about its output has also been referred to as quality estimation.

Confidence Measures is the main subject of 131 publications. 14 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

Confidence measures based on posterior probabilities were proposed by Ueffing et al. (2003); Ueffing and Ney (2007). Other researchers extended this work extend this work using machine learning methods (Blatz et al., 2004) and applied it to multiple, especially non-statistical, systems (Akiba et al., 2004). Sanchis et al. (2007) combine a number of features in a smoothed naive Bayes model. Hardmeier (2011) adds a syntactic matching feature based on tree kernels to a sentence-based confidence estimation measure. Soricut and Echihabi (2010) predict the quality of machine translation output according to a human-understandable metric, the TrustRank. Specia (2011) uses sentence-level confidence estimation to predict post-editing time, automatic metric scores and a subjective assessment of post-editing effort. Related to this, Sánchez-Martínez (2011) train a classifier to determine which machine translation system should be used to translate each sentence.
Confidence measures have been used to for computer aided translation tools (Ueffing and Ney, 2005). Confidence measures have also been used in a self-training method: By translating additional input language text, we bootstrap a parallel corpus to train an additional translation table, but it helps if we filter the bootstrapped corpus using confidence metrics (Ueffing, 2006). This may be also done iteratively, by re-training the model and adding more sentences each time (Ueffing et al., 2007). Alternative, the original training data may be discarded and the phrase table (and other components) only trained from the n-best lists (Chen et al., 2008).
Uchimoto et al. (2005) suggests to use back-translation to assess which input words will cause problems, and then prompt the user of an interactive machine translation system to rephrase the input.

Benchmarks

Discussion

Related Topics

Confidence measures are increasingly of interest for computer aided translation. There have been a few studies on assessing translation difficulty in general, say, for a given language pair. If a human reference translation is given, then evaluation metrics aid in determining the quality of the translation.

New Publications

  • Martins et al. (2017)
  • Jhaveri et al. (2018)
  • Mirkin et al. (2013)
  • Luong et al. (2014)
  • Esplà-Gomis et al. (2015)
  • Logacheva and Specia (2015)
  • Scarton et al. (2015)
  • Kreutzer et al. (2015)
  • Langlois (2015)
  • Logacheva et al. (2015)
  • Esplà-Gomis et al. (2015)
  • Martins and Caseli (2015)
  • Shah et al. (2015)
  • Vela and Tan (2015)
  • Scarton et al. (2015)
  • Shah et al. (2015)
  • Shang et al. (2015)
  • Servan et al. (2015)
  • Kim and Lee (2016)
  • Tezcan et al. (2015)
  • Esplà-Gomis et al. (2016)
  • Kozlova et al. (2016)
  • Martins et al. (2016)
  • Paetzold and Specia (2016)
  • Abdelsalam et al. (2016)
  • Beck et al. (2016)
  • Patel and Sasikumar (2016)
  • Sagemo and Stymne (2016)
  • Tezcan et al. (2016)
  • Logacheva et al. (2016)
  • Scarton et al. (2016)
  • Shah et al. (2016)
  • Bicici et al. (2015)
  • Beck et al. (2016)
  • Sajjad et al. (2016)
  • Shah and Specia (2016)
  • Kim and Lee (2016)
  • Graham et al. (2016)
  • Sperber et al. (2016)
  • Logacheva et al. (2016)
  • Wu et al. (2014)
  • Le et al. (2016)
  • Servan et al. (2015)
  • Shah et al. (2015)
  • Specia et al. (2015)
  • Souza et al. (2015)
  • Graham (2015)
  • Mishra et al. (2013)
  • Avramidis and Popovic (2013)
  • Beck et al. (2013)
  • Bicici (2013)
  • Souza et al. (2013)
  • Scarton and Specia (2014)
  • Kaljahi et al. (2013)
  • Turchi et al. (2014)
  • Souza et al. (2014)
  • Kaljahi et al. (2014)
  • Moreau and Vogel (2014)
  • Kaljahi et al. (2014)
  • Avramidis (2014)
  • Bicici and Way (2014)
  • Scarton and Specia (2014)
  • Hokamp et al. (2014)
  • Beck et al. (2014)
  • Souza et al. (2014)
  • Wisniewski et al. (2014)
  • Cohn and Specia (2013)
  • Turchi et al. (2014)
  • Logacheva and Specia (2014)
  • Luong et al. (2014)
  • Logacheva and Specia (2014)
  • Luong et al. (2014)
  • Specia and Shah (2014)
  • de Souza et al. (2014)
  • Han et al. (2013)
  • Hildebrand and Vogel (2013)
  • Luong et al. (2013)
  • Moreau and Rubino (2013)
  • Rubino et al. (2013)
  • Singh et al. (2013)
  • Turchi et al. (2013)
  • Avramidis (2013)
  • Gispert et al. (2013)
  • Moreau and Vogel (2013)
  • Biçici et al. (2013)
  • Felice and Specia (2013)
  • Specia and Soricut (2013)
  • Wisniewski et al. (2013)
  • Almaghout and Specia (2013)
  • Formiga et al. (2013)
  • Rubino et al. (2013)
  • Shah et al. (2013)
  • González-Rubio et al. (2013)
  • González-Rubio et al. (2013)
  • González-Rubio et al. (2013)
  • González-Rubio et al. (2013)
  • Specia et al. (2011)
  • Specia et al. (2010)
  • Avramidis (2012)
  • Turchi et al. (2012)
  • Okita et al. (2012)
  • Avramidis (2012)
  • Buck (2012)
  • Felice and Specia (2012)
  • Hardmeier et al. (2012)
  • Langlois et al. (2012)
  • Mehdad et al. (2012)
  • Moreau and Vogel (2012)
  • Pighin et al. (2012)
  • Popovic (2012)
  • Rubino et al. (2012)
  • Soricut et al. (2012)
  • Wu and Zhao (2012)
  • Zhuang et al. (2012)
  • Huang (2009)
  • Specia et al. (2009)
  • Rapp (2009)
  • Bernth and Gdaniec (2001)
  • Bach et al. (2011)
  • Ueffing and Ney (2005)
  • Gamon et al. (2005)

Actions

Download

Contribute