Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
文献紹介:Treat the Word As a Whole or Look Inside? ...
Search
Taichi Aida
September 16, 2019
Technology
0
330
文献紹介:Treat the Word As a Whole or Look Inside? Subword Embeddings Model Language Change and Typology
Taichi Aida
September 16, 2019
Tweet
Share
More Decks by Taichi Aida
See All by Taichi Aida
文献紹介:A Multidimensional Framework for Evaluating Lexical Semantic Change with Social Science Applications
a1da4
1
220
YANS2024:目指せ国際会議!「ネットワーキングの極意(国際会議編)」
a1da4
0
100
言語処理学会30周年記念事業留学支援交流会@YANS2024:「学生のための短期留学」
a1da4
1
240
新入生向けチュートリアル:文献のサーベイv2
a1da4
13
8.8k
文献紹介:Isotropic Representation Can Improve Zero-Shot Cross-Lingual Transfer on Multilingual Language Models
a1da4
0
120
文献紹介:WhitenedCSE: Whitening-based Contrastive Learning of Sentence Embeddings
a1da4
1
160
文献紹介:On the Transformation of Latent Space in Fine-Tuned NLP Models
a1da4
0
55
新入生向けチュートリアル:文献のサーベイ
a1da4
0
380
文献紹介:Temporal Attention for Language Models
a1da4
0
290
Other Decks in Technology
See All in Technology
適材適所の技術選定 〜GraphQL・REST API・tRPC〜 / Optimal Technology Selection
kakehashi
1
170
TypeScriptの次なる大進化なるか!? 条件型を返り値とする関数の型推論
uhyo
2
1.6k
The Role of Developer Relations in AI Product Success.
giftojabu1
0
120
安心してください、日本語使えますよ―Ubuntu日本語Remix提供休止に寄せて― 2024-11-17
nobutomurata
1
990
テストコード品質を高めるためにMutation Testingライブラリ・Strykerを実戦導入してみた話
ysknsid25
7
2.6k
SREが投資するAIOps ~ペアーズにおけるLLM for Developerへの取り組み~
takumiogawa
1
100
B2B SaaSから見た最近のC#/.NETの進化
sansantech
PRO
0
720
VideoMamba: State Space Model for Efficient Video Understanding
chou500
0
190
rootlessコンテナのすゝめ - 研究室サーバーでもできる安全なコンテナ管理
kitsuya0828
3
380
ドメイン名の終活について - JPAAWG 7th -
mikit
33
20k
CysharpのOSS群から見るModern C#の現在地
neuecc
2
3.1k
Lambda10周年!Lambdaは何をもたらしたか
smt7174
2
110
Featured
See All Featured
KATA
mclloyd
29
14k
What’s in a name? Adding method to the madness
productmarketing
PRO
22
3.1k
Done Done
chrislema
181
16k
Testing 201, or: Great Expectations
jmmastey
38
7.1k
Optimising Largest Contentful Paint
csswizardry
33
2.9k
Agile that works and the tools we love
rasmusluckow
327
21k
VelocityConf: Rendering Performance Case Studies
addyosmani
325
24k
GitHub's CSS Performance
jonrohan
1030
460k
Speed Design
sergeychernyshev
24
610
[RailsConf 2023 Opening Keynote] The Magic of Rails
eileencodes
28
9.1k
ReactJS: Keep Simple. Everything can be a component!
pedronauck
665
120k
Designing for Performance
lara
604
68k
Transcript
จݙհʢʣ Treat the Word As a Whole or Look Inside?
Subword Embeddings Model Language Change and Typology Yang Xu, Jiasheng Zhang, David Reitter 1st International Workshop on Computational Approaches to Historical Language Change, ACL2019 Ԭٕज़Պֶେֶ ࣗવݴޠॲཧݚڀࣨɹ ૬ాɹଠҰ
Abstract • ݴޠֶతͳԾઆΛௐΔͨΊʹ subword Λߟྀͨ͠୯ޠࢄදݱΛఏҊ • Indo-European ͷݴޠ৽͍͠୯ޠ΄Ͳ subword ʹର͢ΔॏΈ͕૿͑ɺɹ
ٯʹதࠃޠ subword ʹର͢ΔॏΈ͕ݮΓɺ୯ޠʹର͢ΔॏΈ͕૿͑ͨ !2
Motivation w ݴޠֶతͳ݁ ʮதࠃޠʹ͓͍ͯɺ࣌ؒͱͱʹ༏Ґੑ͕୯ԻઅˠೋԻઅʹҠͬͨʯ w Ծઆ ʮݱͷதࠃޠʹ͓͍ͯɺ୯ޠʹؚ·ΕΔࣈจࣈʢTVCXPSEʣ ҙຯతͳׂ͕গͳ͍ʯ !3
Related Work w $#08ʢDPOUFYU ͔ΒUBSHFU Λ༧ଌʣ w $IBSBDUFSFOIBODFEXPSEFNCFEEJOH $8&
w 4LJQHSBNʢUBSHFU ͔ΒDPOUFYU Λ༧ଌʣ w GBTU5FYU vc ui vc ui !4 ୯ޠͱจࣈΛಉ͡ॏཁͰѻ͏
Method w %ZOBNJDTVCXPSEJODPSQPSBUFEFNCFEEJOHNPEFM %4& w %4&$#08 w %4&4( w
୯ޠʹ୯ޠͷॏΈ ͰɺTVCXPSEʹ ͰॏΈ͚͢Δ hw i 1 − hw i !5
Method !6
Experiment w %BUBTFUT w 5SBJOJOHXPSEFNCFEEJOH8JLJQFEJBEBUBCBTFEVNQT w $IJOFTF &OHMJTI 'SFODI (FSNBO
*UBMJBO 4QBOJTI w .PEFM w %4&$#08 %4&4(ʢఏҊख๏ʣ w $8& GBTU5FYU !7
Experiment w ࣮ݧ߲ ͱ୯ޠͷൃੜ࣌ظͱͷ૬ؔ w ൃੜ࣌ظɿ͋Δޠ͕(PPHMF#PPLT/HSBNʹॳΊͯొͨ͠ ޠͷҙຯλεΫ
w &NCFEEJOHͷੑೳΛଌΔ w 4JNJMBSJUZͱ"OBMPHZΛ༻ hw i !8
Result ୯ޠͷॏΈ ͱൃੜ࣌ظͱͷ૬ؔɿ*OEP&VSPQFBOͱதࠃͰਖ਼ର w hw i !9
Result ୯ޠͷॏΈ ͱൃੜ࣌ظͱͷ૬ؔɿ*OEP&VSPQFBOͱதࠃͰਖ਼ର w hw i !10 ͕࣌ਐΉͱ ୯ޠʹର͢ΔॏΈ͕ݮগ ˣ
୯ޠΑΓ 4VCXPSEΛॏࢹ
Result ୯ޠͷॏΈ ͱൃੜ࣌ظͱͷ૬ؔɿ*OEP&VSPQFBOͱதࠃͰਖ਼ର w hw i !11 ͕࣌ਐΉͱ ୯ޠʹର͢ΔॏΈ͕૿Ճ ˣ
4VCXPSEΑΓ ୯ޠΛॏࢹ ʢԾઆ͕͔֬ΊΒΕͨʣ
Result w ͦΕͧΕͷάϧʔϓͰൺֱ w $#08ܥʢ%4&$#08 $8&ʣ w 4LJQHSBNܥʢ%4&4( GBTUUFYUʣ w
%4&4(Ͱੑೳͷ্Λ֬ೝ !12
Conclusion w ԾઆΛݕূ͢ΔҝʹɺTVCXPSEΛߟྀ͢Δ୯ޠࢄදݱΛఏҊͨ͠ w *OEP&VSPQFBOͷݴޠͰ৽͘͠ੜ·ΕΔ୯ޠ΄ͲTVCXPSEʹҙຯͷ ॏΈ͕ॏࢹ͞ΕɺதࠃޠͰٯʹTVCXPSEͷॏΈ͕ݮΓɺ୯ޠͦͷ ͷʹରͯ͠ॏΈ͕ͭ͘Α͏ʹͳͬͨʢԾઆΛݕূͨ͠ʣ !13
None
Discussion w ࣮ݧʹରͯ͠۩ମతͳൺֱΛߦͬͨ w தࠃɿͷۙԽͰٕज़Պֶ͕ൃలͨ͜͠ͱʹΑΓɺ৽͍͠୯ ޠ͕ೖ͖ͬͯͨʁ !15