Conference paper Open Access

Harnessing Indirect Training Data for End-to-End Automatic Speech Translation: Tricks of the Trade

Pino, Juan; Puzon, Liezl; Gu, Jiatao; Ma, Xutai; McCarthy, Arya D.; Gopinath, Deepak


DCAT Export

<?xml version='1.0' encoding='utf-8'?>
<rdf:RDF xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#" xmlns:adms="http://www.w3.org/ns/adms#" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:dct="http://purl.org/dc/terms/" xmlns:dctype="http://purl.org/dc/dcmitype/" xmlns:dcat="http://www.w3.org/ns/dcat#" xmlns:duv="http://www.w3.org/ns/duv#" xmlns:foaf="http://xmlns.com/foaf/0.1/" xmlns:frapo="http://purl.org/cerif/frapo/" xmlns:geo="http://www.w3.org/2003/01/geo/wgs84_pos#" xmlns:gsp="http://www.opengis.net/ont/geosparql#" xmlns:locn="http://www.w3.org/ns/locn#" xmlns:org="http://www.w3.org/ns/org#" xmlns:owl="http://www.w3.org/2002/07/owl#" xmlns:prov="http://www.w3.org/ns/prov#" xmlns:rdfs="http://www.w3.org/2000/01/rdf-schema#" xmlns:schema="http://schema.org/" xmlns:skos="http://www.w3.org/2004/02/skos/core#" xmlns:vcard="http://www.w3.org/2006/vcard/ns#" xmlns:wdrs="http://www.w3.org/2007/05/powder-s#">
  <rdf:Description rdf:about="https://doi.org/10.5281/zenodo.3525032">
    <rdf:type rdf:resource="http://www.w3.org/ns/dcat#Dataset"/>
    <dct:type rdf:resource="http://purl.org/dc/dcmitype/Text"/>
    <dct:identifier rdf:datatype="http://www.w3.org/2001/XMLSchema#anyURI">https://doi.org/10.5281/zenodo.3525032</dct:identifier>
    <foaf:page rdf:resource="https://doi.org/10.5281/zenodo.3525032"/>
    <dct:creator>
      <rdf:Description>
        <rdf:type rdf:resource="http://xmlns.com/foaf/0.1/Agent"/>
        <foaf:name>Pino, Juan</foaf:name>
        <foaf:givenName>Juan</foaf:givenName>
        <foaf:familyName>Pino</foaf:familyName>
        <org:memberOf>
          <foaf:Organization>
            <foaf:name>Facebook</foaf:name>
          </foaf:Organization>
        </org:memberOf>
      </rdf:Description>
    </dct:creator>
    <dct:creator>
      <rdf:Description>
        <rdf:type rdf:resource="http://xmlns.com/foaf/0.1/Agent"/>
        <foaf:name>Puzon, Liezl</foaf:name>
        <foaf:givenName>Liezl</foaf:givenName>
        <foaf:familyName>Puzon</foaf:familyName>
        <org:memberOf>
          <foaf:Organization>
            <foaf:name>Johns Hopkins University</foaf:name>
          </foaf:Organization>
        </org:memberOf>
      </rdf:Description>
    </dct:creator>
    <dct:creator>
      <rdf:Description>
        <rdf:type rdf:resource="http://xmlns.com/foaf/0.1/Agent"/>
        <foaf:name>Gu, Jiatao</foaf:name>
        <foaf:givenName>Jiatao</foaf:givenName>
        <foaf:familyName>Gu</foaf:familyName>
        <org:memberOf>
          <foaf:Organization>
            <foaf:name>Facebook</foaf:name>
          </foaf:Organization>
        </org:memberOf>
      </rdf:Description>
    </dct:creator>
    <dct:creator>
      <rdf:Description>
        <rdf:type rdf:resource="http://xmlns.com/foaf/0.1/Agent"/>
        <foaf:name>Ma, Xutai</foaf:name>
        <foaf:givenName>Xutai</foaf:givenName>
        <foaf:familyName>Ma</foaf:familyName>
        <org:memberOf>
          <foaf:Organization>
            <foaf:name>Facebook &amp; Johns Hopkins University</foaf:name>
          </foaf:Organization>
        </org:memberOf>
      </rdf:Description>
    </dct:creator>
    <dct:creator>
      <rdf:Description>
        <rdf:type rdf:resource="http://xmlns.com/foaf/0.1/Agent"/>
        <foaf:name>McCarthy, Arya D.</foaf:name>
        <foaf:givenName>Arya D.</foaf:givenName>
        <foaf:familyName>McCarthy</foaf:familyName>
        <org:memberOf>
          <foaf:Organization>
            <foaf:name>Facebook &amp; Johns Hopkins University</foaf:name>
          </foaf:Organization>
        </org:memberOf>
      </rdf:Description>
    </dct:creator>
    <dct:creator>
      <rdf:Description>
        <rdf:type rdf:resource="http://xmlns.com/foaf/0.1/Agent"/>
        <foaf:name>Gopinath, Deepak</foaf:name>
        <foaf:givenName>Deepak</foaf:givenName>
        <foaf:familyName>Gopinath</foaf:familyName>
        <org:memberOf>
          <foaf:Organization>
            <foaf:name>Facebook</foaf:name>
          </foaf:Organization>
        </org:memberOf>
      </rdf:Description>
    </dct:creator>
    <dct:title>Harnessing Indirect Training Data for End-to-End Automatic Speech Translation: Tricks of the Trade</dct:title>
    <dct:publisher>
      <foaf:Agent>
        <foaf:name>Zenodo</foaf:name>
      </foaf:Agent>
    </dct:publisher>
    <dct:issued rdf:datatype="http://www.w3.org/2001/XMLSchema#gYear">2019</dct:issued>
    <dct:issued rdf:datatype="http://www.w3.org/2001/XMLSchema#date">2019-11-02</dct:issued>
    <owl:sameAs rdf:resource="https://zenodo.org/record/3525032"/>
    <adms:identifier>
      <adms:Identifier>
        <skos:notation rdf:datatype="http://www.w3.org/2001/XMLSchema#anyURI">https://zenodo.org/record/3525032</skos:notation>
        <adms:schemeAgency>url</adms:schemeAgency>
      </adms:Identifier>
    </adms:identifier>
    <dct:isVersionOf rdf:resource="https://doi.org/10.5281/zenodo.3525031"/>
    <dct:isPartOf rdf:resource="https://zenodo.org/communities/iwslt2019"/>
    <dct:description>&lt;p&gt;For automatic speech translation (AST), end-to-end approaches are outperformed by cascaded models that transcribe with automatic speech recognition (ASR), then trans- late with machine translation (MT). A major cause of the performance gap is that, while existing AST corpora are small, massive datasets exist for both the ASR and MT subsystems. In this work, we evaluate several data augmentation and pretraining approaches for AST, by comparing all on the same datasets. Simple data augmentation by translating ASR transcripts proves most effective on the English&amp;ndash;French augmented LibriSpeech dataset, closing the performance gap from 8.2 to 1.4 BLEU, compared to a very strong cascade that could directly utilize copious ASR and MT data. The same end-to-end approach plus fine-tuning closes the gap on the English&amp;ndash;Romanian MuST-C dataset from 6.7 to 3.7 BLEU. In addition to these results, we present practical rec- ommendations for augmentation and pretraining approaches. Finally, we decrease the performance gap to 0.01 BLEU us- ing a Transformer-based architecture.&lt;/p&gt;</dct:description>
    <dct:accessRights rdf:resource="http://publications.europa.eu/resource/authority/access-right/PUBLIC"/>
    <dct:accessRights>
      <dct:RightsStatement rdf:about="info:eu-repo/semantics/openAccess">
        <rdfs:label>Open Access</rdfs:label>
      </dct:RightsStatement>
    </dct:accessRights>
    <dcat:distribution>
      <dcat:Distribution>
        <dct:license rdf:resource="https://creativecommons.org/licenses/by/4.0/legalcode"/>
        <dcat:accessURL rdf:resource="https://doi.org/10.5281/zenodo.3525032"/>
      </dcat:Distribution>
    </dcat:distribution>
    <dcat:distribution>
      <dcat:Distribution>
        <dcat:accessURL>https://doi.org/10.5281/zenodo.3525032</dcat:accessURL>
        <dcat:byteSize>1586090</dcat:byteSize>
        <dcat:downloadURL>https://zenodo.org/record/3525032/files/IWSLT2019_paper_25.pdf</dcat:downloadURL>
        <dcat:mediaType>application/pdf</dcat:mediaType>
      </dcat:Distribution>
    </dcat:distribution>
  </rdf:Description>
</rdf:RDF>
178
142
views
downloads
All versions This version
Views 178172
Downloads 142142
Data volume 225.2 MB225.2 MB
Unique views 157151
Unique downloads 124124

Share

Cite as