@inproceedings{jin-etal-2021-ji,
title = "基于层间知识蒸馏的神经机器翻译(Inter-layer Knowledge Distillation for Neural Machine Translation)",
author = "Jin, Chang and
Duan, Renchong and
Xiao, Nini and
Duan, Xiangyu",
editor = "Li, Sheng and
Sun, Maosong and
Liu, Yang and
Wu, Hua and
Liu, Kang and
Che, Wanxiang and
He, Shizhu and
Rao, Gaoqi",
booktitle = "Proceedings of the 20th Chinese National Conference on Computational Linguistics",
month = aug,
year = "2021",
address = "Huhhot, China",
publisher = "Chinese Information Processing Society of China",
url = "https://aclanthology.org/2021.ccl-1.16",
pages = "166--175",
abstract = "神经机器翻译(NMT)通常采用多层神经网络模型结构,随着网络层数的加深,所得到的特征也越来越抽象,但是在现有的神经机器翻译模型中,高层的抽象信息仅在预测分布时被利用。为了更好地利用这些信息,本文提出了层间知识蒸馏,目的在于将高层网络的抽象知识迁移到低层网络,使低层网络能够捕捉更加有用的信息,从而提升整个模型的翻译质量。区别于传统教师模型和学生模型的知识蒸馏,层间知识蒸馏实现的是同一个模型内部不同层之间的知识迁移。通过在中文-英语、英语-罗马尼亚语、德语-英语三个数据集上的实验,结果证明层间蒸馏方法能够有效提升翻译性能,分别在中-英、英-罗、德-英上提升1.19,0.72,1.35的BLEU值,同时也证明有效地利用高层信息能够提高神经网络模型的翻译质量。",
language = "Chinese",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="jin-etal-2021-ji">
<titleInfo>
<title>基于层间知识蒸馏的神经机器翻译(Inter-layer Knowledge Distillation for Neural Machine Translation)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Chang</namePart>
<namePart type="family">Jin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Renchong</namePart>
<namePart type="family">Duan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nini</namePart>
<namePart type="family">Xiao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiangyu</namePart>
<namePart type="family">Duan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2021-08</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<language>
<languageTerm type="text">Chinese</languageTerm>
<languageTerm type="code" authority="iso639-2b">chi</languageTerm>
</language>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 20th Chinese National Conference on Computational Linguistics</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sheng</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Maosong</namePart>
<namePart type="family">Sun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yang</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hua</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kang</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wanxiang</namePart>
<namePart type="family">Che</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shizhu</namePart>
<namePart type="family">He</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Gaoqi</namePart>
<namePart type="family">Rao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Chinese Information Processing Society of China</publisher>
<place>
<placeTerm type="text">Huhhot, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>神经机器翻译(NMT)通常采用多层神经网络模型结构,随着网络层数的加深,所得到的特征也越来越抽象,但是在现有的神经机器翻译模型中,高层的抽象信息仅在预测分布时被利用。为了更好地利用这些信息,本文提出了层间知识蒸馏,目的在于将高层网络的抽象知识迁移到低层网络,使低层网络能够捕捉更加有用的信息,从而提升整个模型的翻译质量。区别于传统教师模型和学生模型的知识蒸馏,层间知识蒸馏实现的是同一个模型内部不同层之间的知识迁移。通过在中文-英语、英语-罗马尼亚语、德语-英语三个数据集上的实验,结果证明层间蒸馏方法能够有效提升翻译性能,分别在中-英、英-罗、德-英上提升1.19,0.72,1.35的BLEU值,同时也证明有效地利用高层信息能够提高神经网络模型的翻译质量。</abstract>
<identifier type="citekey">jin-etal-2021-ji</identifier>
<location>
<url>https://aclanthology.org/2021.ccl-1.16</url>
</location>
<part>
<date>2021-08</date>
<extent unit="page">
<start>166</start>
<end>175</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T 基于层间知识蒸馏的神经机器翻译(Inter-layer Knowledge Distillation for Neural Machine Translation)
%A Jin, Chang
%A Duan, Renchong
%A Xiao, Nini
%A Duan, Xiangyu
%Y Li, Sheng
%Y Sun, Maosong
%Y Liu, Yang
%Y Wu, Hua
%Y Liu, Kang
%Y Che, Wanxiang
%Y He, Shizhu
%Y Rao, Gaoqi
%S Proceedings of the 20th Chinese National Conference on Computational Linguistics
%D 2021
%8 August
%I Chinese Information Processing Society of China
%C Huhhot, China
%G Chinese
%F jin-etal-2021-ji
%X 神经机器翻译(NMT)通常采用多层神经网络模型结构,随着网络层数的加深,所得到的特征也越来越抽象,但是在现有的神经机器翻译模型中,高层的抽象信息仅在预测分布时被利用。为了更好地利用这些信息,本文提出了层间知识蒸馏,目的在于将高层网络的抽象知识迁移到低层网络,使低层网络能够捕捉更加有用的信息,从而提升整个模型的翻译质量。区别于传统教师模型和学生模型的知识蒸馏,层间知识蒸馏实现的是同一个模型内部不同层之间的知识迁移。通过在中文-英语、英语-罗马尼亚语、德语-英语三个数据集上的实验,结果证明层间蒸馏方法能够有效提升翻译性能,分别在中-英、英-罗、德-英上提升1.19,0.72,1.35的BLEU值,同时也证明有效地利用高层信息能够提高神经网络模型的翻译质量。
%U https://aclanthology.org/2021.ccl-1.16
%P 166-175
Markdown (Informal)
[基于层间知识蒸馏的神经机器翻译(Inter-layer Knowledge Distillation for Neural Machine Translation)](https://aclanthology.org/2021.ccl-1.16) (Jin et al., CCL 2021)
ACL