Lego-MT: Learning Detachable Models for Massively Multilingual Machine Translation
Multilingual neural machine translation (MNMT) aims to build a unified model for many language directions. Existing monolithic models for MNMT encounter two challenges: parameter interference among languages and inefficient inference for large models. In this paper, we revisit the classic multi-way...
Gespeichert in:
Veröffentlicht in: | arXiv.org 2023-07 |
---|---|
Hauptverfasser: | , , , , , , |
Format: | Artikel |
Sprache: | eng |
Schlagworte: | |
Online-Zugang: | Volltext |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
container_end_page | |
---|---|
container_issue | |
container_start_page | |
container_title | arXiv.org |
container_volume | |
creator | Yuan, Fei Lu, Yinquan Zhu, WenHao Kong, Lingpeng Li, Lei Yu, Qiao Xu, Jingjing |
description | Multilingual neural machine translation (MNMT) aims to build a unified model for many language directions. Existing monolithic models for MNMT encounter two challenges: parameter interference among languages and inefficient inference for large models. In this paper, we revisit the classic multi-way structures and develop a detachable model by assigning each language (or group of languages) to an individual branch that supports plug-and-play training and inference. To address the needs of learning representations for all languages in a unified space, we propose a novel efficient training recipe, upon which we build an effective detachable model, Lego-MT. For a fair comparison, we collect data from OPUS and build a translation benchmark covering 433 languages and 1.3B parallel data. Experiments show that Lego-MT with 1.2B parameters brings an average gain of 3.2 spBLEU. It even outperforms M2M-100 with 12B parameters. The proposed training recipe brings a 28.2\(\times\) speedup over the conventional multi-way training method.\footnote{ \url{https://github.com/CONE-MT/Lego-MT}.} |
format | Article |
fullrecord | <record><control><sourceid>proquest</sourceid><recordid>TN_cdi_proquest_journals_2756547212</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>2756547212</sourcerecordid><originalsourceid>FETCH-proquest_journals_27565472123</originalsourceid><addsrcrecordid>eNqNir0KwjAYAIMgWLTvEHAutEnTiqs_ODSLdJdYv7YpIdF8ieDb28EHcDq4uwVJGOdFtisZW5EUccrznFU1E4In5NrA4DLZ7mkDylttB3qEoLpR3Q1Q6R5gkPbOU6kQ9RvMh8pogjbzGZWZdTdqC7T1yqJRQTu7IcteGYT0xzXZnk_t4ZI9vXtFwHCbXPR2TjdWi0qUNSsY_-_6AqGpP2g</addsrcrecordid><sourcetype>Aggregation Database</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype><pqid>2756547212</pqid></control><display><type>article</type><title>Lego-MT: Learning Detachable Models for Massively Multilingual Machine Translation</title><source>Free E- Journals</source><creator>Yuan, Fei ; Lu, Yinquan ; Zhu, WenHao ; Kong, Lingpeng ; Li, Lei ; Yu, Qiao ; Xu, Jingjing</creator><creatorcontrib>Yuan, Fei ; Lu, Yinquan ; Zhu, WenHao ; Kong, Lingpeng ; Li, Lei ; Yu, Qiao ; Xu, Jingjing</creatorcontrib><description>Multilingual neural machine translation (MNMT) aims to build a unified model for many language directions. Existing monolithic models for MNMT encounter two challenges: parameter interference among languages and inefficient inference for large models. In this paper, we revisit the classic multi-way structures and develop a detachable model by assigning each language (or group of languages) to an individual branch that supports plug-and-play training and inference. To address the needs of learning representations for all languages in a unified space, we propose a novel efficient training recipe, upon which we build an effective detachable model, Lego-MT. For a fair comparison, we collect data from OPUS and build a translation benchmark covering 433 languages and 1.3B parallel data. Experiments show that Lego-MT with 1.2B parameters brings an average gain of 3.2 spBLEU. It even outperforms M2M-100 with 12B parameters. The proposed training recipe brings a 28.2\(\times\) speedup over the conventional multi-way training method.\footnote{ \url{https://github.com/CONE-MT/Lego-MT}.}</description><identifier>EISSN: 2331-8422</identifier><language>eng</language><publisher>Ithaca: Cornell University Library, arXiv.org</publisher><subject>Data collection ; Machine translation ; Multilingualism ; Training</subject><ispartof>arXiv.org, 2023-07</ispartof><rights>2023. This work is published under http://creativecommons.org/licenses/by/4.0/ (the “License”). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.</rights><oa>free_for_read</oa><woscitedreferencessubscribed>false</woscitedreferencessubscribed></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><link.rule.ids>780,784</link.rule.ids></links><search><creatorcontrib>Yuan, Fei</creatorcontrib><creatorcontrib>Lu, Yinquan</creatorcontrib><creatorcontrib>Zhu, WenHao</creatorcontrib><creatorcontrib>Kong, Lingpeng</creatorcontrib><creatorcontrib>Li, Lei</creatorcontrib><creatorcontrib>Yu, Qiao</creatorcontrib><creatorcontrib>Xu, Jingjing</creatorcontrib><title>Lego-MT: Learning Detachable Models for Massively Multilingual Machine Translation</title><title>arXiv.org</title><description>Multilingual neural machine translation (MNMT) aims to build a unified model for many language directions. Existing monolithic models for MNMT encounter two challenges: parameter interference among languages and inefficient inference for large models. In this paper, we revisit the classic multi-way structures and develop a detachable model by assigning each language (or group of languages) to an individual branch that supports plug-and-play training and inference. To address the needs of learning representations for all languages in a unified space, we propose a novel efficient training recipe, upon which we build an effective detachable model, Lego-MT. For a fair comparison, we collect data from OPUS and build a translation benchmark covering 433 languages and 1.3B parallel data. Experiments show that Lego-MT with 1.2B parameters brings an average gain of 3.2 spBLEU. It even outperforms M2M-100 with 12B parameters. The proposed training recipe brings a 28.2\(\times\) speedup over the conventional multi-way training method.\footnote{ \url{https://github.com/CONE-MT/Lego-MT}.}</description><subject>Data collection</subject><subject>Machine translation</subject><subject>Multilingualism</subject><subject>Training</subject><issn>2331-8422</issn><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2023</creationdate><recordtype>article</recordtype><sourceid>ABUWG</sourceid><sourceid>AFKRA</sourceid><sourceid>AZQEC</sourceid><sourceid>BENPR</sourceid><sourceid>CCPQU</sourceid><sourceid>DWQXO</sourceid><recordid>eNqNir0KwjAYAIMgWLTvEHAutEnTiqs_ODSLdJdYv7YpIdF8ieDb28EHcDq4uwVJGOdFtisZW5EUccrznFU1E4In5NrA4DLZ7mkDylttB3qEoLpR3Q1Q6R5gkPbOU6kQ9RvMh8pogjbzGZWZdTdqC7T1yqJRQTu7IcteGYT0xzXZnk_t4ZI9vXtFwHCbXPR2TjdWi0qUNSsY_-_6AqGpP2g</recordid><startdate>20230719</startdate><enddate>20230719</enddate><creator>Yuan, Fei</creator><creator>Lu, Yinquan</creator><creator>Zhu, WenHao</creator><creator>Kong, Lingpeng</creator><creator>Li, Lei</creator><creator>Yu, Qiao</creator><creator>Xu, Jingjing</creator><general>Cornell University Library, arXiv.org</general><scope>8FE</scope><scope>8FG</scope><scope>ABJCF</scope><scope>ABUWG</scope><scope>AFKRA</scope><scope>AZQEC</scope><scope>BENPR</scope><scope>BGLVJ</scope><scope>CCPQU</scope><scope>DWQXO</scope><scope>HCIFZ</scope><scope>L6V</scope><scope>M7S</scope><scope>PIMPY</scope><scope>PQEST</scope><scope>PQQKQ</scope><scope>PQUKI</scope><scope>PRINS</scope><scope>PTHSS</scope></search><sort><creationdate>20230719</creationdate><title>Lego-MT: Learning Detachable Models for Massively Multilingual Machine Translation</title><author>Yuan, Fei ; Lu, Yinquan ; Zhu, WenHao ; Kong, Lingpeng ; Li, Lei ; Yu, Qiao ; Xu, Jingjing</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-proquest_journals_27565472123</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2023</creationdate><topic>Data collection</topic><topic>Machine translation</topic><topic>Multilingualism</topic><topic>Training</topic><toplevel>online_resources</toplevel><creatorcontrib>Yuan, Fei</creatorcontrib><creatorcontrib>Lu, Yinquan</creatorcontrib><creatorcontrib>Zhu, WenHao</creatorcontrib><creatorcontrib>Kong, Lingpeng</creatorcontrib><creatorcontrib>Li, Lei</creatorcontrib><creatorcontrib>Yu, Qiao</creatorcontrib><creatorcontrib>Xu, Jingjing</creatorcontrib><collection>ProQuest SciTech Collection</collection><collection>ProQuest Technology Collection</collection><collection>Materials Science & Engineering Collection</collection><collection>ProQuest Central (Alumni Edition)</collection><collection>ProQuest Central UK/Ireland</collection><collection>ProQuest Central Essentials</collection><collection>ProQuest Central</collection><collection>Technology Collection</collection><collection>ProQuest One Community College</collection><collection>ProQuest Central Korea</collection><collection>SciTech Premium Collection</collection><collection>ProQuest Engineering Collection</collection><collection>Engineering Database</collection><collection>Publicly Available Content Database</collection><collection>ProQuest One Academic Eastern Edition (DO NOT USE)</collection><collection>ProQuest One Academic</collection><collection>ProQuest One Academic UKI Edition</collection><collection>ProQuest Central China</collection><collection>Engineering Collection</collection></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext</fulltext></delivery><addata><au>Yuan, Fei</au><au>Lu, Yinquan</au><au>Zhu, WenHao</au><au>Kong, Lingpeng</au><au>Li, Lei</au><au>Yu, Qiao</au><au>Xu, Jingjing</au><format>book</format><genre>document</genre><ristype>GEN</ristype><atitle>Lego-MT: Learning Detachable Models for Massively Multilingual Machine Translation</atitle><jtitle>arXiv.org</jtitle><date>2023-07-19</date><risdate>2023</risdate><eissn>2331-8422</eissn><abstract>Multilingual neural machine translation (MNMT) aims to build a unified model for many language directions. Existing monolithic models for MNMT encounter two challenges: parameter interference among languages and inefficient inference for large models. In this paper, we revisit the classic multi-way structures and develop a detachable model by assigning each language (or group of languages) to an individual branch that supports plug-and-play training and inference. To address the needs of learning representations for all languages in a unified space, we propose a novel efficient training recipe, upon which we build an effective detachable model, Lego-MT. For a fair comparison, we collect data from OPUS and build a translation benchmark covering 433 languages and 1.3B parallel data. Experiments show that Lego-MT with 1.2B parameters brings an average gain of 3.2 spBLEU. It even outperforms M2M-100 with 12B parameters. The proposed training recipe brings a 28.2\(\times\) speedup over the conventional multi-way training method.\footnote{ \url{https://github.com/CONE-MT/Lego-MT}.}</abstract><cop>Ithaca</cop><pub>Cornell University Library, arXiv.org</pub><oa>free_for_read</oa></addata></record> |
fulltext | fulltext |
identifier | EISSN: 2331-8422 |
ispartof | arXiv.org, 2023-07 |
issn | 2331-8422 |
language | eng |
recordid | cdi_proquest_journals_2756547212 |
source | Free E- Journals |
subjects | Data collection Machine translation Multilingualism Training |
title | Lego-MT: Learning Detachable Models for Massively Multilingual Machine Translation |
url | https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2025-01-12T12%3A13%3A11IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-proquest&rft_val_fmt=info:ofi/fmt:kev:mtx:book&rft.genre=document&rft.atitle=Lego-MT:%20Learning%20Detachable%20Models%20for%20Massively%20Multilingual%20Machine%20Translation&rft.jtitle=arXiv.org&rft.au=Yuan,%20Fei&rft.date=2023-07-19&rft.eissn=2331-8422&rft_id=info:doi/&rft_dat=%3Cproquest%3E2756547212%3C/proquest%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_pqid=2756547212&rft_id=info:pmid/&rfr_iscdi=true |