mycoreobjectrosdok_document_0000027951rosdok_document_0000027951rosdokdocumentrosdok_derivate_00002298622026-10-13T13:21:01.452Z2026-10-10T09:49:36.678ZeditorKEeditorKEpublished1rosdok_derivate_0000229862truetruederivate_types:fulltextVolltextfulltextwf_edit_epub wf_register_epubstate:publishedstate:publishedveröffentlichtpublishedrosdok_id000058201986180786Oau2026-10-102026-10-13T13:21:00ZrdaConverted from PICA to MODS using Pica2MODS XSLT Transformer 2.14-SNAPSHOT [SCM: "" "" ""] with mode 'DEFAULT'.WorkingpaperAnapproach for output decoding of neural networksWe review two decoding methods for recurrent neural networks – based on the well known weighted Levenshtein distance and the connectionist temporal classification (CTC) – and we introduce a new method – the dynamic Levenstein distance (DynWL). We compare these three methods analytically in terms of time complexity and error performance. Although the approaches are different, there are deep connections between These ways of decoding. Finally, we test on the Arabic and French ICDAR data sets. Our experiments show that CTC yields the smallest error rates. Nevertheless, there are scenarios where DynWL is a good choice between performance and time complexity.TobiasStrauß1984 -VerfasserInaut11362945540000-0002-5344-4188RogerLabahn1959 -VerfasserInaut1709996450000-0003-1901-9644GundramLeifertVerfasserInautWelfWustlichVerfasserInauthttps://purl.uni-rostock.de/rosdok/id00005820urn:nbn:de:gbv:28-rosdok_id00005820-510.18453/rosdok_id00005820510 MathematikMathematisch-Naturwissenschaftliche Fakultätalle Rechte vorbehaltenRechteinhaber Universität Rostockfrei zugänglich (Open Access)Lizenz Metadaten: CC0en1 Online-Ressource (7 Seiten)2012University of Rostock, Institute of MathematicsRostockmonographic20122026Universitätsbibliothek RostockRostock2026Universitätsbibliothek Rostockhttps://purl.uni-rostock.de/rosdok/id00005820Tobias Strauß, Roger Labahn, Gundram Leifert, Welf Wustlichrosdok_id000058161986139832Odvhttps://purl.uni-rostock.de/rosdok/id00005816Preprints aus dem Institut für Mathematik3255797-812/30012-0003doctype:epubdoctype:epubDokumenttypDocument typedoctype:epub.workingpaperdoctype:epub.workingpaperWorkingpaperworking paperdiniPublType:workingPaper diniPublType2022:WorkingPaperinfo:eu-repo/semantics/workingPaperdocumentdiniPublType2022:WorkingPaperdiniPublType2022:WorkingPaperArbeitspapierWorking paperKDSF (Pu37)SDNB:510SDNB:510510 Mathematik510 Mathematicsinstitution:unirostockinstitution:unirostockUniversität RostockUniversity of RostockUniversität RostockUniversität RostockUni.Rostockhttp://d-nb.info/gnd/38329-6institution:unirostock.mnfinstitution:unirostock.mnfMathematisch-Naturwissenschaftliche FakultätFaculty of Mathematics and Natural SciencesUniversität Rostock. Mathematisch-Naturwissenschaftliche FakultätMathematisch-Natur-<br />wissenschaftliche FakultätUni.Rostock.Fakultaet.MNFhttp://d-nb.info/gnd/2147083-2licenseinfo:worklicenseinfo:workWerkworklicenseinfo:work.rightsreservedlicenseinfo:work.rightsreservedalle Rechte vorbehaltenall rights reserved/creativecommons/r/reserved/0.9/88x31.png[DE-28]Urheberrechtsschutz 1.0$gRights Statements$uhttp://rightsstatements.org/vocab/InC/1.0/http://rightsstatements.org/vocab/InC/1.0/http://rightsstatements.org/vocab/InC/1.0/licenseinfo:depositlicenseinfo:depositVeröffentlichungsgenehmigungpermission to storelicenseinfo:deposit.rightsunirostocklicenseinfo:deposit.rightsunirostockRechteinhaber Universität Rostockrightsholder Rostock Universityaccesscondition:openaccessaccesscondition:openaccessfrei zugänglich (Open Access)open accesshttp://purl.org/coar/access_right/c_abf2OAfreeinfo:eu-repo/semantics/openAccess[DE-28]Open Access$gControlled Vocabulary for Access Rights$uhttp://purl.org/coar/access_right/c_abf2licenseinfo:metadatalicenseinfo:metadataLizenzen für Metadatenlicenseinfo:metadata.cc0licenseinfo:metadata.cc0Lizenz Metadaten: CC0license metadata: CC0/creativecommons/p/zero/1.0/88x31.pnghttps://creativecommons.org/publicdomain/zero/1.0/rfc5646:enrfc5646:enEnglischEnglishengengAnapproach for output decoding of neural networksAn approach for output decoding of neural networksgnd:1136294554orcid:0000-0002-5344-4188gnd:170999645orcid:0000-0003-1901-9644gnd:1136294554orcid:0000-0002-5344-4188gnd:170999645orcid:0000-0003-1901-9644rosdok_document_0000027951-d4134118e50gnd:1136294554orcid:0000-0002-5344-4188Tobias StraußTobias Straußrosdok_document_0000027951-d4134118e66gnd:170999645orcid:0000-0003-1901-9644Roger LabahnRoger Labahnrosdok_document_0000027951-d4134118e82Gundram LeifertGundram Leifertrosdok_document_0000027951-d4134118e92Welf WustlichWelf WustlichTobias StraußRoger LabahnGundram LeifertWelf WustlichTobias StraußRoger LabahnGundram LeifertWelf WustlichTobias StraußRoger LabahnGundram LeifertWelf WustlichRostockUniversity of Rostock, Institute of Mathematicsepub.workingpaperhttps://purl.uni-rostock.de/rosdok/id00005820urn:nbn:de:gbv:28-rosdok_id00005820-510.18453/rosdok_id00005820We review two decoding methods for recurrent neural networks – based on the well known weighted Levenshtein distance and the connectionist temporal classification (CTC) – and we introduce a new method – the dynamic Levenstein distance (DynWL). We compare these three methods analytically in terms of time complexity and error performance. Although the approaches are different, there are deep connections between These ways of decoding. Finally, we test on the Arabic and French ICDAR data sets. Our experiments show that CTC yields the smallest error rates. Nevertheless, there are scenarios where DynWL is a good choice between performance and time complexity.20122012Preprints aus dem Institut für Mathematikrosdok_id000058161986139832Odvhttps://purl.uni-rostock.de/rosdok/id00005816Preprints aus dem Institut für Mathematik3255797-812/30012-0003https://purl.uni-rostock.de/rosdok/id000058163255797-8Tobias Strauß, Roger Labahn, Gundram Leifert, Welf Wustlichepub.workingpaper1 Strauss_approach_2012.pdf 171020 10a6aa03bd7824832ec8558852b94e2a rosdok_id000058201986180786Oau2026-10-102026-10-13T13:21:00ZrdaConverted from PICA to MODS using Pica2MODS XSLT Transformer 2.14-SNAPSHOT [SCM: "" "" ""] with mode 'DEFAULT'.WorkingpaperAnapproach for output decoding of neural networksWe review two decoding methods for recurrent neural networks – based on the well known weighted Levenshtein distance and the connectionist temporal classification (CTC) – and we introduce a new method – the dynamic Levenstein distance (DynWL). We compare these three methods analytically in terms of time complexity and error performance. Although the approaches are different, there are deep connections between These ways of decoding. Finally, we test on the Arabic and French ICDAR data sets. Our experiments show that CTC yields the smallest error rates. Nevertheless, there are scenarios where DynWL is a good choice between performance and time complexity.TobiasStrauß1984 -VerfasserInaut11362945540000-0002-5344-4188RogerLabahn1959 -VerfasserInaut1709996450000-0003-1901-9644GundramLeifertVerfasserInautWelfWustlichVerfasserInauthttps://purl.uni-rostock.de/rosdok/id00005820urn:nbn:de:gbv:28-rosdok_id00005820-510.18453/rosdok_id00005820510 MathematikMathematisch-Naturwissenschaftliche Fakultätalle Rechte vorbehaltenRechteinhaber Universität Rostockfrei zugänglich (Open Access)Lizenz Metadaten: CC0en1 Online-Ressource (7 Seiten)2012University of Rostock, Institute of MathematicsRostockmonographic20122026Universitätsbibliothek RostockRostock2026Universitätsbibliothek Rostockhttps://purl.uni-rostock.de/rosdok/id00005820Tobias Strauß, Roger Labahn, Gundram Leifert, Welf Wustlichrosdok_id000058161986139832Odvhttps://purl.uni-rostock.de/rosdok/id00005816Preprints aus dem Institut für Mathematik3255797-812/30012-0003 2026-10-10T09:49:36.678Z 2026-10-13T13:21:01.452Z 2026-10-23T13:21:01.461Z editorKE {"identifier":"rosdok/id00005820","type":"local_id","additional":"","service":"MCRLocalID","created":"2026-10-10T09:49:36.687Z"} {"identifier":"urn:nbn:de:gbv:28-rosdok_id00005820-5","type":"dnbUrn","additional":"","service":"RosDokURN","created":"2026-10-10T09:49:36.694Z"} {"identifier":"https://purl.uni-rostock.de/rosdok/id00005820","type":"purl","additional":"","service":"RosDokPURL","created":"2026-10-10T09:49:36.697Z","registrationStarted":"2026-10-10T09:53:32.175Z"} {"identifier":"10.18453/rosdok_id00005820","type":"doi","additional":"","service":"RosDokDOI","created":"2026-10-10T09:49:36.700Z","registrationStarted":"2026-10-10T09:53:31.924Z","registered":"2026-10-10T09:53:33.416Z"} editorKEfulltextfile/rosdok_document_0000027951/rosdok_derivate_0000229862/Strauss_approach_2012.pdfAnapproach for output decoding of neural networksAn approach for output decoding of neural networksgnd:1136294554orcid:0000-0002-5344-4188gnd:170999645orcid:0000-0003-1901-9644gnd:1136294554orcid:0000-0002-5344-4188gnd:170999645orcid:0000-0003-1901-9644rosdok_document_0000027951-d4134118e50gnd:1136294554orcid:0000-0002-5344-4188Tobias StraußTobias Straußrosdok_document_0000027951-d4134118e66gnd:170999645orcid:0000-0003-1901-9644Roger LabahnRoger Labahnrosdok_document_0000027951-d4134118e82Gundram LeifertGundram Leifertrosdok_document_0000027951-d4134118e92Welf WustlichWelf WustlichTobias StraußRoger LabahnGundram LeifertWelf WustlichTobias StraußRoger LabahnGundram LeifertWelf WustlichTobias StraußRoger LabahnGundram LeifertWelf WustlichRostockUniversity of Rostock, Institute of Mathematicsepub.workingpaperhttps://purl.uni-rostock.de/rosdok/id00005820urn:nbn:de:gbv:28-rosdok_id00005820-510.18453/rosdok_id00005820We review two decoding methods for recurrent neural networks – based on the well known weighted Levenshtein distance and the connectionist temporal classification (CTC) – and we introduce a new method – the dynamic Levenstein distance (DynWL). We compare these three methods analytically in terms of time complexity and error performance. Although the approaches are different, there are deep connections between These ways of decoding. Finally, we test on the Arabic and French ICDAR data sets. Our experiments show that CTC yields the smallest error rates. Nevertheless, there are scenarios where DynWL is a good choice between performance and time complexity.20122012Preprints aus dem Institut für Mathematikrosdok_id000058161986139832Odvhttps://purl.uni-rostock.de/rosdok/id00005816Preprints aus dem Institut für Mathematik3255797-812/30012-0003https://purl.uni-rostock.de/rosdok/id000058163255797-8Tobias Strauß, Roger Labahn, Gundram Leifert, Welf Wustlich[xslt]Saxonrosdok_id00005820https://purl.uni-rostock.de/rosdok/id00005820198618078610.18453/rosdok_id00005820urn:nbn:de:gbv:28-rosdok_id00005820-5Tobias Strauß, Roger Labahn, Gundram Leifert, Welf WustlichStrauß Tobias, Labahn Roger, Leifert Gundram, Wustlich WelfAn approach for output decoding of neural networksWorkingpaperworking paperUniversity of Rostock, Institute of Mathematics, 2012We review two decoding methods for recurrent neural networks – based on the well known weighted Levenshtein distance and the connectionist temporal classification (CTC) – and we introduce a new method – the dynamic Levenstein distance (DynWL). We compare these three methods analytically in terms of…rosdok_id000058160012-000312/3Tobias StraußRoger LabahnGundram LeifertWelf WustlichAnapproach for output decoding of neural networksPreprints aus dem Institut für MathematikUniversitätsbibliothek Rostockhttps://purl.uni-rostock.de/rosdok/id00005820TobiasStrauß1984 -VerfasserInaut11362945540000-0002-5344-4188RogerLabahn1959 -VerfasserInaut1709996450000-0003-1901-9644GundramLeifertVerfasserInautWelfWustlichVerfasserInaut[purl]https://purl.uni-rostock.de/rosdok/id00005820[urn]urn:nbn:de:gbv:28-rosdok_id00005820-5[doi]10.18453/rosdok_id00005820open_access20122012epoch:21th_centuryrfc5646:endoctype:epub.workingpaperaccesscondition:openaccessSDNB:510institution:unirostock.mnfstate:published