"This sentence is wrong." Detecting errors in machine-translated sentences
Identifieur interne : 000132 ( PascalFrancis/Corpus ); précédent : 000131; suivant : 000133"This sentence is wrong." Detecting errors in machine-translated sentences
Auteurs : Sylvain Raybaud ; David Langlois ; Kamel SmaïliSource :
- Machine translation [ 0922-6567 ] ; 2011.
Descripteurs français
- Pascal (Inist)
English descriptors
- KwdEn :
Abstract
Machine translation systems are not reliable enough to be used "as is": except for the most simple tasks, they can only be used to grasp the general meaning of a text or assist human translators. The purpose of confidence measures is to detect erroneous words or sentences produced by a machine translation system. In this article, after reviewing the mathematical foundations of confidence estimation, we propose a comparison of several state-of-the-art confidence measures, predictive parameters and classifiers. We also propose two original confidence measures based on Mutual Information and a method for automatically generating data for training and testing classifiers. We applied these techniques to data from the WMT campaign 2008 and found that the best confidence measures yielded an Equal Error Rate of 36.3% at word level and 34.2% at sentence level, but combining different measures reduced these rates to 35.0% and 29.0%, respectively. We also present the results of an experiment aimed at determining how helpful confidence measures are in a post-editing task. Preliminary results suggest that our system is not yet ready to efficiently help post-editors, but we now have both software and a protocol that we can apply to further experiments, and user feedback has indicated aspects which must be improved in order to increase the level of helpfulness of confidence measures.
Notice en format standard (ISO 2709)
Pour connaître la documentation sur le format Inist Standard.
pA |
|
---|
Format Inist (serveur)
NO : | FRANCIS 12-0189218 INIST |
---|---|
ET : | "This sentence is wrong." Detecting errors in machine-translated sentences |
AU : | RAYBAUD (Sylvain); LANGLOIS (David); SMAÏLI (Kamel) |
AF : | PAROLE, LORIA, BP 239/54506 Nancy/France (1 aut., 2 aut., 3 aut.) |
DT : | Publication en série; Niveau analytique |
SO : | Machine translation; ISSN 0922-6567; Allemagne; Da. 2011; Vol. 25; No. 1; Pp. 1-34; Bibl. 1 p.1/4 |
LA : | Anglais |
EA : | Machine translation systems are not reliable enough to be used "as is": except for the most simple tasks, they can only be used to grasp the general meaning of a text or assist human translators. The purpose of confidence measures is to detect erroneous words or sentences produced by a machine translation system. In this article, after reviewing the mathematical foundations of confidence estimation, we propose a comparison of several state-of-the-art confidence measures, predictive parameters and classifiers. We also propose two original confidence measures based on Mutual Information and a method for automatically generating data for training and testing classifiers. We applied these techniques to data from the WMT campaign 2008 and found that the best confidence measures yielded an Equal Error Rate of 36.3% at word level and 34.2% at sentence level, but combining different measures reduced these rates to 35.0% and 29.0%, respectively. We also present the results of an experiment aimed at determining how helpful confidence measures are in a post-editing task. Preliminary results suggest that our system is not yet ready to efficiently help post-editors, but we now have both software and a protocol that we can apply to further experiments, and user feedback has indicated aspects which must be improved in order to increase the level of helpfulness of confidence measures. |
CC : | 52477; 524 |
FD : | Traduction automatique; Classificateur; Linguistique informatique; Traitement automatique des langues naturelles; Erreur; Evaluation; Réseau neuronal; Mathématiques; Machine vecteur support |
ED : | Machine translation; Classifier; Computational linguistics; Natural language processing; Error; Assessment; Neural network; Mathematics |
LO : | INIST-21070.354000508919860010 |
ID : | 12-0189218 |
Links to Exploration step
Francis:12-0189218Le document en format XML
<record><TEI><teiHeader><fileDesc><titleStmt><title xml:lang="en" level="a">"This sentence is wrong." Detecting errors in machine-translated sentences</title>
<author><name sortKey="Raybaud, Sylvain" sort="Raybaud, Sylvain" uniqKey="Raybaud S" first="Sylvain" last="Raybaud">Sylvain Raybaud</name>
<affiliation><inist:fA14 i1="01"><s1>PAROLE, LORIA, BP 239</s1>
<s2>54506 Nancy</s2>
<s3>FRA</s3>
<sZ>1 aut.</sZ>
<sZ>2 aut.</sZ>
<sZ>3 aut.</sZ>
</inist:fA14>
</affiliation>
</author>
<author><name sortKey="Langlois, David" sort="Langlois, David" uniqKey="Langlois D" first="David" last="Langlois">David Langlois</name>
<affiliation><inist:fA14 i1="01"><s1>PAROLE, LORIA, BP 239</s1>
<s2>54506 Nancy</s2>
<s3>FRA</s3>
<sZ>1 aut.</sZ>
<sZ>2 aut.</sZ>
<sZ>3 aut.</sZ>
</inist:fA14>
</affiliation>
</author>
<author><name sortKey="Smaili, Kamel" sort="Smaili, Kamel" uniqKey="Smaili K" first="Kamel" last="Smaïli">Kamel Smaïli</name>
<affiliation><inist:fA14 i1="01"><s1>PAROLE, LORIA, BP 239</s1>
<s2>54506 Nancy</s2>
<s3>FRA</s3>
<sZ>1 aut.</sZ>
<sZ>2 aut.</sZ>
<sZ>3 aut.</sZ>
</inist:fA14>
</affiliation>
</author>
</titleStmt>
<publicationStmt><idno type="wicri:source">INIST</idno>
<idno type="inist">12-0189218</idno>
<date when="2011">2011</date>
<idno type="stanalyst">FRANCIS 12-0189218 INIST</idno>
<idno type="RBID">Francis:12-0189218</idno>
<idno type="wicri:Area/PascalFrancis/Corpus">000132</idno>
</publicationStmt>
<sourceDesc><biblStruct><analytic><title xml:lang="en" level="a">"This sentence is wrong." Detecting errors in machine-translated sentences</title>
<author><name sortKey="Raybaud, Sylvain" sort="Raybaud, Sylvain" uniqKey="Raybaud S" first="Sylvain" last="Raybaud">Sylvain Raybaud</name>
<affiliation><inist:fA14 i1="01"><s1>PAROLE, LORIA, BP 239</s1>
<s2>54506 Nancy</s2>
<s3>FRA</s3>
<sZ>1 aut.</sZ>
<sZ>2 aut.</sZ>
<sZ>3 aut.</sZ>
</inist:fA14>
</affiliation>
</author>
<author><name sortKey="Langlois, David" sort="Langlois, David" uniqKey="Langlois D" first="David" last="Langlois">David Langlois</name>
<affiliation><inist:fA14 i1="01"><s1>PAROLE, LORIA, BP 239</s1>
<s2>54506 Nancy</s2>
<s3>FRA</s3>
<sZ>1 aut.</sZ>
<sZ>2 aut.</sZ>
<sZ>3 aut.</sZ>
</inist:fA14>
</affiliation>
</author>
<author><name sortKey="Smaili, Kamel" sort="Smaili, Kamel" uniqKey="Smaili K" first="Kamel" last="Smaïli">Kamel Smaïli</name>
<affiliation><inist:fA14 i1="01"><s1>PAROLE, LORIA, BP 239</s1>
<s2>54506 Nancy</s2>
<s3>FRA</s3>
<sZ>1 aut.</sZ>
<sZ>2 aut.</sZ>
<sZ>3 aut.</sZ>
</inist:fA14>
</affiliation>
</author>
</analytic>
<series><title level="j" type="main">Machine translation</title>
<title level="j" type="abbreviated">Mach. transl.</title>
<idno type="ISSN">0922-6567</idno>
<imprint><date when="2011">2011</date>
</imprint>
</series>
</biblStruct>
</sourceDesc>
<seriesStmt><title level="j" type="main">Machine translation</title>
<title level="j" type="abbreviated">Mach. transl.</title>
<idno type="ISSN">0922-6567</idno>
</seriesStmt>
</fileDesc>
<profileDesc><textClass><keywords scheme="KwdEn" xml:lang="en"><term>Assessment</term>
<term>Classifier</term>
<term>Computational linguistics</term>
<term>Error</term>
<term>Machine translation</term>
<term>Mathematics</term>
<term>Natural language processing</term>
<term>Neural network</term>
</keywords>
<keywords scheme="Pascal" xml:lang="fr"><term>Traduction automatique</term>
<term>Classificateur</term>
<term>Linguistique informatique</term>
<term>Traitement automatique des langues naturelles</term>
<term>Erreur</term>
<term>Evaluation</term>
<term>Réseau neuronal</term>
<term>Mathématiques</term>
<term>Machine vecteur support</term>
</keywords>
</textClass>
</profileDesc>
</teiHeader>
<front><div type="abstract" xml:lang="en">Machine translation systems are not reliable enough to be used "as is": except for the most simple tasks, they can only be used to grasp the general meaning of a text or assist human translators. The purpose of confidence measures is to detect erroneous words or sentences produced by a machine translation system. In this article, after reviewing the mathematical foundations of confidence estimation, we propose a comparison of several state-of-the-art confidence measures, predictive parameters and classifiers. We also propose two original confidence measures based on Mutual Information and a method for automatically generating data for training and testing classifiers. We applied these techniques to data from the WMT campaign 2008 and found that the best confidence measures yielded an Equal Error Rate of 36.3% at word level and 34.2% at sentence level, but combining different measures reduced these rates to 35.0% and 29.0%, respectively. We also present the results of an experiment aimed at determining how helpful confidence measures are in a post-editing task. Preliminary results suggest that our system is not yet ready to efficiently help post-editors, but we now have both software and a protocol that we can apply to further experiments, and user feedback has indicated aspects which must be improved in order to increase the level of helpfulness of confidence measures.</div>
</front>
</TEI>
<inist><standard h6="B"><pA><fA01 i1="01" i2="1"><s0>0922-6567</s0>
</fA01>
<fA03 i2="1"><s0>Mach. transl.</s0>
</fA03>
<fA05><s2>25</s2>
</fA05>
<fA06><s2>1</s2>
</fA06>
<fA08 i1="01" i2="1" l="ENG"><s1>"This sentence is wrong." Detecting errors in machine-translated sentences</s1>
</fA08>
<fA11 i1="01" i2="1"><s1>RAYBAUD (Sylvain)</s1>
</fA11>
<fA11 i1="02" i2="1"><s1>LANGLOIS (David)</s1>
</fA11>
<fA11 i1="03" i2="1"><s1>SMAÏLI (Kamel)</s1>
</fA11>
<fA14 i1="01"><s1>PAROLE, LORIA, BP 239</s1>
<s2>54506 Nancy</s2>
<s3>FRA</s3>
<sZ>1 aut.</sZ>
<sZ>2 aut.</sZ>
<sZ>3 aut.</sZ>
</fA14>
<fA20><s1>1-34</s1>
</fA20>
<fA21><s1>2011</s1>
</fA21>
<fA23 i1="01"><s0>ENG</s0>
</fA23>
<fA43 i1="01"><s1>INIST</s1>
<s2>21070</s2>
<s5>354000508919860010</s5>
</fA43>
<fA44><s0>0000</s0>
<s1>© 2012 INIST-CNRS. All rights reserved.</s1>
</fA44>
<fA45><s0>1 p.1/4</s0>
</fA45>
<fA47 i1="01" i2="1"><s0>12-0189218</s0>
</fA47>
<fA60><s1>P</s1>
</fA60>
<fA61><s0>A</s0>
</fA61>
<fA64 i1="01" i2="1"><s0>Machine translation</s0>
</fA64>
<fA66 i1="01"><s0>DEU</s0>
</fA66>
<fC01 i1="01" l="ENG"><s0>Machine translation systems are not reliable enough to be used "as is": except for the most simple tasks, they can only be used to grasp the general meaning of a text or assist human translators. The purpose of confidence measures is to detect erroneous words or sentences produced by a machine translation system. In this article, after reviewing the mathematical foundations of confidence estimation, we propose a comparison of several state-of-the-art confidence measures, predictive parameters and classifiers. We also propose two original confidence measures based on Mutual Information and a method for automatically generating data for training and testing classifiers. We applied these techniques to data from the WMT campaign 2008 and found that the best confidence measures yielded an Equal Error Rate of 36.3% at word level and 34.2% at sentence level, but combining different measures reduced these rates to 35.0% and 29.0%, respectively. We also present the results of an experiment aimed at determining how helpful confidence measures are in a post-editing task. Preliminary results suggest that our system is not yet ready to efficiently help post-editors, but we now have both software and a protocol that we can apply to further experiments, and user feedback has indicated aspects which must be improved in order to increase the level of helpfulness of confidence measures.</s0>
</fC01>
<fC02 i1="01" i2="L"><s0>52477</s0>
<s1>XV</s1>
</fC02>
<fC02 i1="02" i2="L"><s0>524</s0>
</fC02>
<fC03 i1="01" i2="L" l="FRE"><s0>Traduction automatique</s0>
<s2>NI</s2>
<s5>01</s5>
</fC03>
<fC03 i1="01" i2="L" l="ENG"><s0>Machine translation</s0>
<s2>NI</s2>
<s5>01</s5>
</fC03>
<fC03 i1="02" i2="L" l="FRE"><s0>Classificateur</s0>
<s2>NI</s2>
<s5>02</s5>
</fC03>
<fC03 i1="02" i2="L" l="ENG"><s0>Classifier</s0>
<s2>NI</s2>
<s5>02</s5>
</fC03>
<fC03 i1="03" i2="L" l="FRE"><s0>Linguistique informatique</s0>
<s2>NI</s2>
<s5>03</s5>
</fC03>
<fC03 i1="03" i2="L" l="ENG"><s0>Computational linguistics</s0>
<s2>NI</s2>
<s5>03</s5>
</fC03>
<fC03 i1="04" i2="L" l="FRE"><s0>Traitement automatique des langues naturelles</s0>
<s2>NI</s2>
<s5>04</s5>
</fC03>
<fC03 i1="04" i2="L" l="ENG"><s0>Natural language processing</s0>
<s2>NI</s2>
<s5>04</s5>
</fC03>
<fC03 i1="05" i2="L" l="FRE"><s0>Erreur</s0>
<s2>NI</s2>
<s5>05</s5>
</fC03>
<fC03 i1="05" i2="L" l="ENG"><s0>Error</s0>
<s2>NI</s2>
<s5>05</s5>
</fC03>
<fC03 i1="06" i2="L" l="FRE"><s0>Evaluation</s0>
<s2>NI</s2>
<s5>06</s5>
</fC03>
<fC03 i1="06" i2="L" l="ENG"><s0>Assessment</s0>
<s2>NI</s2>
<s5>06</s5>
</fC03>
<fC03 i1="07" i2="L" l="FRE"><s0>Réseau neuronal</s0>
<s2>NI</s2>
<s5>07</s5>
</fC03>
<fC03 i1="07" i2="L" l="ENG"><s0>Neural network</s0>
<s2>NI</s2>
<s5>07</s5>
</fC03>
<fC03 i1="08" i2="L" l="FRE"><s0>Mathématiques</s0>
<s2>NI</s2>
<s5>08</s5>
</fC03>
<fC03 i1="08" i2="L" l="ENG"><s0>Mathematics</s0>
<s2>NI</s2>
<s5>08</s5>
</fC03>
<fC03 i1="09" i2="L" l="FRE"><s0>Machine vecteur support</s0>
<s4>INC</s4>
<s5>31</s5>
</fC03>
<fN21><s1>149</s1>
</fN21>
</pA>
</standard>
<server><NO>FRANCIS 12-0189218 INIST</NO>
<ET>"This sentence is wrong." Detecting errors in machine-translated sentences</ET>
<AU>RAYBAUD (Sylvain); LANGLOIS (David); SMAÏLI (Kamel)</AU>
<AF>PAROLE, LORIA, BP 239/54506 Nancy/France (1 aut., 2 aut., 3 aut.)</AF>
<DT>Publication en série; Niveau analytique</DT>
<SO>Machine translation; ISSN 0922-6567; Allemagne; Da. 2011; Vol. 25; No. 1; Pp. 1-34; Bibl. 1 p.1/4</SO>
<LA>Anglais</LA>
<EA>Machine translation systems are not reliable enough to be used "as is": except for the most simple tasks, they can only be used to grasp the general meaning of a text or assist human translators. The purpose of confidence measures is to detect erroneous words or sentences produced by a machine translation system. In this article, after reviewing the mathematical foundations of confidence estimation, we propose a comparison of several state-of-the-art confidence measures, predictive parameters and classifiers. We also propose two original confidence measures based on Mutual Information and a method for automatically generating data for training and testing classifiers. We applied these techniques to data from the WMT campaign 2008 and found that the best confidence measures yielded an Equal Error Rate of 36.3% at word level and 34.2% at sentence level, but combining different measures reduced these rates to 35.0% and 29.0%, respectively. We also present the results of an experiment aimed at determining how helpful confidence measures are in a post-editing task. Preliminary results suggest that our system is not yet ready to efficiently help post-editors, but we now have both software and a protocol that we can apply to further experiments, and user feedback has indicated aspects which must be improved in order to increase the level of helpfulness of confidence measures.</EA>
<CC>52477; 524</CC>
<FD>Traduction automatique; Classificateur; Linguistique informatique; Traitement automatique des langues naturelles; Erreur; Evaluation; Réseau neuronal; Mathématiques; Machine vecteur support</FD>
<ED>Machine translation; Classifier; Computational linguistics; Natural language processing; Error; Assessment; Neural network; Mathematics</ED>
<LO>INIST-21070.354000508919860010</LO>
<ID>12-0189218</ID>
</server>
</inist>
</record>
Pour manipuler ce document sous Unix (Dilib)
EXPLOR_STEP=$WICRI_ROOT/Wicri/Lorraine/explor/InforLorV4/Data/PascalFrancis/Corpus
HfdSelect -h $EXPLOR_STEP/biblio.hfd -nk 000132 | SxmlIndent | more
Ou
HfdSelect -h $EXPLOR_AREA/Data/PascalFrancis/Corpus/biblio.hfd -nk 000132 | SxmlIndent | more
Pour mettre un lien sur cette page dans le réseau Wicri
{{Explor lien |wiki= Wicri/Lorraine |area= InforLorV4 |flux= PascalFrancis |étape= Corpus |type= RBID |clé= Francis:12-0189218 |texte= "This sentence is wrong." Detecting errors in machine-translated sentences }}
![]() | This area was generated with Dilib version V0.6.33. | ![]() |