<?xml version="1.0" encoding="utf-8"?>
<TEI xmlns="http://www.tei-c.org/ns/1.0" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xmlns:hal="http://hal.archives-ouvertes.fr/" xmlns:gml="http://www.opengis.net/gml/3.3/" xmlns:gmlce="http://www.opengis.net/gml/3.3/ce" version="1.1" xsi:schemaLocation="http://www.tei-c.org/ns/1.0 http://api.archives-ouvertes.fr/documents/aofr-sword.xsd">
  <teiHeader>
    <fileDesc>
      <titleStmt>
        <title>HAL TEI export of hal-04678036</title>
      </titleStmt>
      <publicationStmt>
        <distributor>CCSD</distributor>
        <availability status="restricted">
          <licence target="https://creativecommons.org/publicdomain/zero/1.0/">CC0 1.0 - Universal</licence>
        </availability>
        <date when="2026-05-18T15:26:34+02:00"/>
      </publicationStmt>
      <sourceDesc>
        <p part="N">HAL API Platform</p>
      </sourceDesc>
    </fileDesc>
  </teiHeader>
  <text>
    <body>
      <listBibl>
        <biblFull>
          <titleStmt>
            <title xml:lang="en">Analysis of Tokenization Metrics and their Correlation with Multilingual Machine Translation Performance</title>
            <title xml:lang="fr">Analyse des métriques de tokenisation et leur corrélation avec les performances de traduction automatique multilingue</title>
            <author role="aut">
              <persName>
                <forename type="first">Serge</forename>
                <surname>Molina</surname>
              </persName>
              <email type="md5">8d29d1ce85557281e4eae5e6f13a5582</email>
              <email type="domain">irit.fr</email>
              <idno type="idhal" notation="string">sergemolinairit</idno>
              <idno type="idhal" notation="numeric">1419333</idno>
              <idno type="halauthorid" notation="string">1178707-1419333</idno>
              <idno type="ORCID">https://orcid.org/0009-0007-7894-7041</idno>
              <affiliation ref="#struct-34499"/>
              <affiliation ref="#struct-217752"/>
            </author>
            <author role="aut">
              <persName>
                <forename type="first">Josiane</forename>
                <surname>Mothe</surname>
              </persName>
              <email type="md5">657fe4ee30d17b8313a2ae629951ac8e</email>
              <email type="domain">irit.fr</email>
              <idno type="idhal" notation="string">josianemothe</idno>
              <idno type="idhal" notation="numeric">735149</idno>
              <idno type="halauthorid" notation="string">6914-735149</idno>
              <idno type="ORCID">https://orcid.org/0000-0001-9273-2193</idno>
              <idno type="IDREF">https://www.idref.fr/087097222</idno>
              <affiliation ref="#struct-402019"/>
              <affiliation ref="#struct-116256"/>
            </author>
            <editor role="depositor">
              <persName>
                <forename>CCSD</forename>
                <surname>Sciencesconf.org</surname>
              </persName>
              <email type="md5">2dac992f64ee924a7ad555e53daad36f</email>
              <email type="domain">sciencesconf.org</email>
            </editor>
          </titleStmt>
          <editionStmt>
            <edition n="v1" type="current">
              <date type="whenSubmitted">2024-08-26 16:16:44</date>
              <date type="whenModified">2025-10-22 18:04:09</date>
              <date type="whenReleased">2024-08-26 16:18:45</date>
              <date type="whenProduced">2024-07-08</date>
              <date type="whenEndEmbargoed">2024-08-26</date>
              <ref type="file" target="https://hal.science/hal-04678036v1/document">
                <date notBefore="2024-08-26"/>
              </ref>
              <ref type="file" subtype="author" n="1" target="https://hal.science/hal-04678036v1/file/JEP_TALN_RECITAL_2024_2.pdf" id="file-4678036-4064953">
                <date notBefore="2024-08-26"/>
              </ref>
            </edition>
            <respStmt>
              <resp>contributor</resp>
              <name key="167428">
                <persName>
                  <forename>CCSD</forename>
                  <surname>Sciencesconf.org</surname>
                </persName>
                <email type="md5">2dac992f64ee924a7ad555e53daad36f</email>
                <email type="domain">sciencesconf.org</email>
              </name>
            </respStmt>
          </editionStmt>
          <publicationStmt>
            <distributor>CCSD</distributor>
            <idno type="halId">hal-04678036</idno>
            <idno type="halUri">https://hal.science/hal-04678036</idno>
            <idno type="halBibtex">molina:hal-04678036</idno>
            <idno type="halRefHtml">&lt;i&gt;EvalLLM2024 - Atelier sur l'évaluation des modèles génératifs et challence d'extraction d'information few-shot (Atelier Eval LLM 2024) @ TALN 2024,&lt;/i&gt;, AMIAD, Jul 2024, Toulouse, France</idno>
            <idno type="halRef">EvalLLM2024 - Atelier sur l'évaluation des modèles génératifs et challence d'extraction d'information few-shot (Atelier Eval LLM 2024) @ TALN 2024,, AMIAD, Jul 2024, Toulouse, France</idno>
            <availability status="restricted">
              <licence target="https://about.hal.science/hal-authorisation-v1/">HAL Authorization<ref corresp="#file-4678036-4064953"/></licence>
            </availability>
          </publicationStmt>
          <seriesStmt>
            <idno type="stamp" n="UNIV-TLSE2">Université Toulouse 2</idno>
            <idno type="stamp" n="UNIV-TLSE3">Université de Toulouse</idno>
            <idno type="stamp" n="CNRS">CNRS - Centre national de la recherche scientifique</idno>
            <idno type="stamp" n="UT1-CAPITOLE">Université Toulouse 1 Capitole</idno>
            <idno type="stamp" n="IRIT">IRIT - Institut de Recherche en Informatique de Toulouse</idno>
            <idno type="stamp" n="IRIT-SIG" corresp="IRIT">Systèmes d’Informations Généralisées</idno>
            <idno type="stamp" n="IRIT-GD" corresp="IRIT">Gestion des Données</idno>
            <idno type="stamp" n="IRIT-UT2J">IRIT - Université Toulouse Jean Jaurès</idno>
            <idno type="stamp" n="IRIT-UT3">IRIT - Université Toulouse III Paul Sabatier</idno>
            <idno type="stamp" n="TOULOUSE-INP">Toulouse - INP</idno>
            <idno type="stamp" n="UNIV-UT3">Université Toulouse 3</idno>
            <idno type="stamp" n="UT3-INP">Université de Toulouse / Toulouse INP</idno>
            <idno type="stamp" n="UT3-TOULOUSEINP">Université de Toulouse / Toulouse INP</idno>
            <idno type="stamp" n="EVALLLM2024">Institut des sciences informatiques et de leurs interactions - CNRS Sciences informatiques [INS2I-CNRS]</idno>
          </seriesStmt>
          <notesStmt>
            <note type="audience" n="2">International</note>
            <note type="invited" n="0">No</note>
            <note type="popular" n="0">No</note>
            <note type="peer" n="1">Yes</note>
            <note type="proceedings" n="1">Yes</note>
          </notesStmt>
          <sourceDesc>
            <biblStruct>
              <analytic>
                <title xml:lang="en">Analysis of Tokenization Metrics and their Correlation with Multilingual Machine Translation Performance</title>
                <title xml:lang="fr">Analyse des métriques de tokenisation et leur corrélation avec les performances de traduction automatique multilingue</title>
                <author role="aut">
                  <persName>
                    <forename type="first">Serge</forename>
                    <surname>Molina</surname>
                  </persName>
                  <email type="md5">8d29d1ce85557281e4eae5e6f13a5582</email>
                  <email type="domain">irit.fr</email>
                  <idno type="idhal" notation="string">sergemolinairit</idno>
                  <idno type="idhal" notation="numeric">1419333</idno>
                  <idno type="halauthorid" notation="string">1178707-1419333</idno>
                  <idno type="ORCID">https://orcid.org/0009-0007-7894-7041</idno>
                  <affiliation ref="#struct-34499"/>
                  <affiliation ref="#struct-217752"/>
                </author>
                <author role="aut">
                  <persName>
                    <forename type="first">Josiane</forename>
                    <surname>Mothe</surname>
                  </persName>
                  <email type="md5">657fe4ee30d17b8313a2ae629951ac8e</email>
                  <email type="domain">irit.fr</email>
                  <idno type="idhal" notation="string">josianemothe</idno>
                  <idno type="idhal" notation="numeric">735149</idno>
                  <idno type="halauthorid" notation="string">6914-735149</idno>
                  <idno type="ORCID">https://orcid.org/0000-0001-9273-2193</idno>
                  <idno type="IDREF">https://www.idref.fr/087097222</idno>
                  <affiliation ref="#struct-402019"/>
                  <affiliation ref="#struct-116256"/>
                </author>
              </analytic>
              <monogr>
                <meeting>
                  <title>EvalLLM2024 - Atelier sur l'évaluation des modèles génératifs et challence d'extraction d'information few-shot (Atelier Eval LLM 2024) @ TALN 2024,</title>
                  <date type="start">2024-07-08</date>
                  <date type="end">2024-07-08</date>
                  <settlement>Toulouse</settlement>
                  <country key="FR">France</country>
                </meeting>
                <respStmt>
                  <resp>conferenceOrganizer</resp>
                  <name>AMIAD</name>
                </respStmt>
                <imprint/>
              </monogr>
              <ref target="https://evalllm2024.sciencesconf.org/569547" type="seeAlso"/>
            </biblStruct>
          </sourceDesc>
          <profileDesc>
            <langUsage>
              <language ident="fr">French</language>
            </langUsage>
            <textClass>
              <keywords scheme="author">
                <term xml:lang="en">BERTScore</term>
                <term xml:lang="en">chrF</term>
                <term xml:lang="en">BLEU</term>
                <term xml:lang="en">Pretrained Language Models</term>
                <term xml:lang="en">FLORES-200</term>
                <term xml:lang="en">Byte Pair Encoding</term>
                <term xml:lang="en">Tokenization</term>
                <term xml:lang="en">Tokenizer</term>
                <term xml:lang="en">Neural Machine Translation</term>
                <term xml:lang="en">Multilingual Machine Translation</term>
                <term xml:lang="en">Natural Language Processing</term>
                <term xml:lang="en">SentencePiece</term>
                <term xml:lang="fr">BERTScore</term>
                <term xml:lang="fr">chrF</term>
                <term xml:lang="fr">BLEU</term>
                <term xml:lang="fr">Modèles de Langue Pré-entraînés</term>
                <term xml:lang="fr">FLORES-200</term>
                <term xml:lang="fr">BPE</term>
                <term xml:lang="fr">SentencePiece</term>
                <term xml:lang="fr">Tokenisation</term>
                <term xml:lang="fr">Tokeniseur</term>
                <term xml:lang="fr">Traduc- tion Automatique Neuronale</term>
                <term xml:lang="fr">Traduction Automatique Multilingue</term>
                <term xml:lang="fr">Traitement Automatique des Langues</term>
              </keywords>
              <classCode scheme="halDomain" n="info">Computer Science [cs]</classCode>
              <classCode scheme="halTypology" n="COMM">Conference papers</classCode>
              <classCode scheme="halOldTypology" n="COMM">Conference papers</classCode>
              <classCode scheme="halTreeTypology" n="COMM">Conference papers</classCode>
            </textClass>
            <abstract xml:lang="en">
              <p>This article studies the tokenization outcomes produced by two subword-based tokenizers in thecontext of multilingual machine translation using the FLORES-200 corpus. It focuses on 9 languages from 3 families, each with varying levels of digital presence. First, we evaluate translation performance between language and family pairs. Then, we explore the potential relation between measures characterizing the tokenization results and translation performance. We aim to validate or invalidate the hypothesis that the difficulty a multilingual translation model would have in translating a text could be inferred from features related to the tokens produced by its tokenizer. The measures are calculated based on the source sentence and its expert-produced reference translated version. We could not detect strong correlation between the features related to tokens and the translation effectiveness. This preliminary study suggests however directions for future research aiming to include more qualitative features to evaluate the impact of tokenized representations on translation performance.</p>
            </abstract>
            <abstract xml:lang="fr">
              <p>Cet article étudie le résultat de la tokenisation obtenu par deux tokeniseurs basés sur des sous-mots dans le contexte de la traduction automatique multilingue en utilisant le corpus FLORES-200. Il s’intéresse à 9 langues issues de 3 familles, avec différents niveaux de présence numérique. Tout d’abord nous évaluons les performances de traductions entre paires de langues et de familles de langue. Ensuite nous explorons les liens potentiels entre des mesures de caractérisation des résultats de la tokenisation et les performances de traduction. Notre objectif est de confirmer ou réfuter l’hypothèse selon laquelle la capacité d’un modèle de traduction multilingue de type encodeur-décodeur à traduire un texte peut être prédite en examinant les caractéristiques des tokens générés. Nous nous appuyons sur des mesures calculées à partir de la phrase source et de sa version traduite par un expert (version de référence), représentées sous forme de chaînes de caractères et de listes de tokens générés par le tokeniseur du modèle considéré. Nous n’avons pas détecté de corrélation fortes entre entre les caractéristiques des tokens et l’efficacité de traduction. Cette étude préliminaire propose toutefois des pistes pour de futures recherches visant à intégrer des mesures de caractérisation plus qualitativespour évaluer l’impact des représentations tokenisées sur les performances de traduction.</p>
            </abstract>
          </profileDesc>
        </biblFull>
      </listBibl>
    </body>
    <back>
      <listOrg type="structures">
        <org type="laboratory" xml:id="struct-34499" status="OLD">
          <idno type="IdRef">030947448</idno>
          <idno type="ISNI">0000000121927225</idno>
          <idno type="RNSR">199511949P</idno>
          <idno type="ROR">https://ror.org/01rx4qw44</idno>
          <orgName>Institut de recherche en informatique de Toulouse</orgName>
          <orgName type="acronym">IRIT</orgName>
          <date type="start">1990-01-01</date>
          <date type="end">2025-01-01</date>
          <desc>
            <address>
              <addrLine>118 Route de Narbonne, F-31062 Toulouse Cedex 9</addrLine>
              <country key="FR"/>
            </address>
            <ref type="url">http://www.irit.fr/</ref>
          </desc>
          <listRelation>
            <relation active="#struct-81148" type="direct"/>
            <relation name="443875" active="#struct-443875" type="indirect"/>
            <relation active="#struct-116256" type="direct"/>
            <relation active="#struct-443875" type="indirect"/>
            <relation active="#struct-217752" type="direct"/>
            <relation name="UMR5505 / URA1399" active="#struct-441569" type="direct"/>
            <relation active="#struct-448187" type="direct"/>
            <relation active="#struct-1149079" type="direct"/>
            <relation name="FED4171" active="#struct-217752" type="direct"/>
          </listRelation>
        </org>
        <org type="institution" xml:id="struct-217752" status="OLD">
          <idno type="IdRef">026404672</idno>
          <idno type="ISNI">0000000121617331</idno>
          <idno type="ROR">https://ror.org/02v6kpv12</idno>
          <idno type="Wikidata">Q1273188</idno>
          <orgName>Université Toulouse III - Paul Sabatier</orgName>
          <orgName type="acronym">UT3</orgName>
          <date type="end">2025-01-01</date>
          <desc>
            <address>
              <addrLine>118 route de Narbonne - 31062 Toulouse</addrLine>
              <country key="FR"/>
            </address>
            <ref type="url">http://www.univ-tlse3.fr/</ref>
          </desc>
          <listRelation>
            <relation active="#struct-443875" type="direct"/>
          </listRelation>
        </org>
        <org type="researchteam" xml:id="struct-402019" status="OLD">
          <orgName>Systèmes d’Informations Généralisées</orgName>
          <orgName type="acronym">IRIT-SIG</orgName>
          <date type="end">2025-01-01</date>
          <desc>
            <address>
              <addrLine>IRIT118 Route de Narbonne31062 Toulouse Cedex 9</addrLine>
              <country key="FR"/>
            </address>
            <ref type="url">https://www.irit.fr/departement/gestion-des-donnees/equipe-sig/</ref>
          </desc>
          <listRelation>
            <relation active="#struct-34499" type="direct"/>
            <relation active="#struct-81148" type="indirect"/>
            <relation name="443875" active="#struct-443875" type="indirect"/>
            <relation active="#struct-116256" type="indirect"/>
            <relation active="#struct-443875" type="indirect"/>
            <relation active="#struct-217752" type="indirect"/>
            <relation name="UMR5505 / URA1399" active="#struct-441569" type="indirect"/>
            <relation active="#struct-448187" type="indirect"/>
            <relation active="#struct-1149079" type="indirect"/>
            <relation name="FED4171" active="#struct-217752" type="indirect"/>
          </listRelation>
        </org>
        <org type="institution" xml:id="struct-116256" status="VALID">
          <idno type="IdRef">026403994</idno>
          <idno type="ROR">https://ror.org/04ezk3x31</idno>
          <orgName>Université Toulouse - Jean Jaurès</orgName>
          <orgName type="acronym">UT2J</orgName>
          <date type="start">1971-01-01</date>
          <desc>
            <address>
              <addrLine>5 allées Antonio Machado - 31058 Toulouse Cedex 9</addrLine>
              <country key="FR"/>
            </address>
            <ref type="url">http://www.univ-tlse2.fr</ref>
          </desc>
          <listRelation>
            <relation active="#struct-443875" type="direct"/>
          </listRelation>
        </org>
        <org type="institution" xml:id="struct-81148" status="VALID">
          <idno type="IdRef">026404354</idno>
          <idno type="ISNI">0000000121902837</idno>
          <idno type="ROR">https://ror.org/0443n9e75</idno>
          <orgName>Université Toulouse Capitole</orgName>
          <orgName type="acronym">UT Capitole</orgName>
          <date type="start">1970-01-01</date>
          <desc>
            <address>
              <addrLine>2 rue du Doyen-Gabriel-Marty - 31042 Toulouse Cedex 9</addrLine>
              <country key="FR"/>
            </address>
            <ref type="url">http://www.ut-capitole.fr/</ref>
          </desc>
          <listRelation>
            <relation name="443875" active="#struct-443875" type="direct"/>
          </listRelation>
        </org>
        <org type="regroupinstitution" xml:id="struct-443875" status="VALID">
          <idno type="ROR">https://ror.org/017tgbk05</idno>
          <orgName>Communauté d'universités et établissements de Toulouse</orgName>
          <orgName type="acronym">Comue de Toulouse</orgName>
          <desc>
            <address>
              <addrLine>41 Allée Jules Guesde, 31000 Toulouse</addrLine>
              <country key="FR"/>
            </address>
            <ref type="url">https://www.univ-toulouse.fr/</ref>
          </desc>
        </org>
        <org type="regroupinstitution" xml:id="struct-441569" status="VALID">
          <idno type="IdRef">02636817X</idno>
          <idno type="ISNI">0000000122597504</idno>
          <idno type="ROR">https://ror.org/02feahw73</idno>
          <orgName>Centre National de la Recherche Scientifique</orgName>
          <orgName type="acronym">CNRS</orgName>
          <date type="start">1939-10-19</date>
          <desc>
            <address>
              <country key="FR"/>
            </address>
            <ref type="url">https://www.cnrs.fr/</ref>
          </desc>
        </org>
        <org type="institution" xml:id="struct-448187" status="VALID">
          <idno type="IdRef">026388820</idno>
          <idno type="ISNI">0000000121766169</idno>
          <idno type="ROR">https://ror.org/033p9g875</idno>
          <orgName>Institut National Polytechnique (Toulouse)</orgName>
          <orgName type="acronym">Toulouse INP</orgName>
          <desc>
            <address>
              <country key="FR"/>
            </address>
            <ref type="url">https://www.inp-toulouse.fr/</ref>
          </desc>
          <listRelation>
            <relation active="#struct-443875" type="direct"/>
          </listRelation>
        </org>
        <org type="regrouplaboratory" xml:id="struct-1149079" status="OLD">
          <orgName>Toulouse Mind &amp; Brain Institut</orgName>
          <orgName type="acronym">TMBI</orgName>
          <date type="start">2021-01-01</date>
          <date type="end">2025-01-01</date>
          <desc>
            <address>
              <addrLine>CHU Purpan, Pavillon Baudot, 31059 Toulouse cedex 9</addrLine>
              <country key="FR"/>
            </address>
            <ref type="url">https://www.tmbi.fr/</ref>
          </desc>
          <listRelation>
            <relation active="#struct-116256" type="direct"/>
            <relation active="#struct-443875" type="indirect"/>
            <relation name="FED4171" active="#struct-217752" type="direct"/>
          </listRelation>
        </org>
      </listOrg>
    </back>
  </text>
</TEI>