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:Auxiliary Objectives for Neural Error Dete...
Search
Atsushi
December 21, 2018
Technology
0
92
文献紹介:Auxiliary Objectives for Neural Error Detection Models
2018/12/21 文献紹介
長岡技術科学大学
自然言語処理研究室
Atsushi
December 21, 2018
Tweet
Share
More Decks by Atsushi
See All by Atsushi
文献紹介:Automated Evaluation of Out-of-Context Errors
atsumikan
0
99
文献紹介:Correction of OCR Word Segmentation Errors in Articles from the ACL Collection through Neural Machine Translation Methods
atsumikan
0
160
文献紹介:Wronging a Right: Generating Better Errors to Improve Grammatical Error Detection
atsumikan
0
120
文献紹介:Low-resource OCR error detection and correction in French Clinical Texts
atsumikan
0
130
文献紹介:CMMC-BDRC Solution to the NLP-TEA-2018 Chinese Grammatical Error Diagnosis Task
atsumikan
0
130
文献紹介 : Fluency Boost Learning and Inference for Neural Grammatical Error Correction
atsumikan
0
180
文献紹介:語彙の概念化と Wikipediaを用いた英字略語の意味推定方法
atsumikan
0
150
文献紹介:The Effect of Error Rate in Artificially Generated Data for Automatic Preposition and Determiner Correction
atsumikan
0
140
文献紹介: Automatic Annotation and Evaluation of Error Types for Grammatical Error Correction
atsumikan
0
170
Other Decks in Technology
See All in Technology
フルカイテン株式会社 エンジニア向け採用資料
fullkaiten
0
10k
AWS re:Inventre:cap ~AmazonNova 2 Omniのワークショップを体験してきた~
nrinetcom
PRO
0
130
複雑さを受け入れるか、拒むか? - 事業成長とともに育ったモノリスを前に私が考えたこと #RSGT2026
murabayashi
1
1.3k
「リリースファースト」の実感を届けるには 〜停滞するチームに変化を起こすアプローチ〜 #RSGT2026
kintotechdev
0
580
製造業から学んだ「本質を守り現場に合わせるアジャイル実践」
kamitokusari
0
330
AI との良い付き合い方を僕らは誰も知らない (WSS 2026 静岡版)
asei
1
230
歴史から学ぶ、Goのメモリ管理基礎
logica0419
10
2.1k
Master Dataグループ紹介資料
sansan33
PRO
1
4.2k
2025年 山梨の技術コミュニティを振り返る
yuukis
0
150
BidiAgent と Nova 2 Sonic から考える音声 AI について
yama3133
2
150
業務の煩悩を祓うAI活用術108選 / AI 108 Usages
smartbank
9
19k
Introduction to Bill One Development Engineer
sansan33
PRO
0
340
Featured
See All Featured
Visualization
eitanlees
150
16k
Become a Pro
speakerdeck
PRO
31
5.8k
Money Talks: Using Revenue to Get Sh*t Done
nikkihalliwell
0
130
Designing for Performance
lara
610
70k
Building an army of robots
kneath
306
46k
Lightning talk: Run Django tests with GitHub Actions
sabderemane
0
97
Context Engineering - Making Every Token Count
addyosmani
9
590
For a Future-Friendly Web
brad_frost
180
10k
Fireside Chat
paigeccino
41
3.8k
A Guide to Academic Writing Using Generative AI - A Workshop
ks91
PRO
0
170
Practical Tips for Bootstrapping Information Extraction Pipelines
honnibal
25
1.7k
Rebuilding a faster, lazier Slack
samanthasiow
85
9.3k
Transcript
Ԭٕज़Պֶେֶࣗવݴޠॲཧݚڀࣨ ੁ३ࢤ จݙհ ݄ .BSFL3FJ )FMFO:BOOBLPVEBLJT 1SPDFFEJOHTPGUIFUI8PSLTIPQPO*OOPWBUJWF6TFPG/-1GPS #VJMEJOH&EVDBUJPOBM"QQMJDBUJPOT QBHFTr$PQFOIBHFO
%FONBSL 4FQUFNCFS "VYJMJBSZ0CKFDUJWFT GPS/FVSBM&SSPS%FUFDUJPO.PEFMT
"CTUSBDU w ݴޠֶशऀͷ࡞ͨ͠จͷޡΓݕग़ w χϡʔϥϧΛ༻͍ͨܥྻϥϕϦϯάख๏ʹ͓͍ͯɺ ิॿతͷ܇࿅ͷ༗༻ੑΛݕূ w ҎલͷޡΓݕγεςϜͱಉ͡ͷύϥϝʔλͰ ΑΓྑ͍ύϑΥʔϚϯεΛୡ !2
*OUSPEVDUJPO w ݴޠֶशऀͷ࡞ͨ͠จͰ༷ʑͳλΠϓͷޡΓΛ ࣝผ͢Δඞཁ͕͋Δ ɾػೳޠͷޡͬͨ༻ ɾ༰ޠͷҙຯతͳޡΓʢલஔࢺɾܗ༰ࢺͷΈ߹ΘͤͳͲʣ w ݴޠͷΑΓྑ͍දݱΛֶͼɺจ຺ʹ͓͚ΔޡΓΛ ΑΓਖ਼֬ʹݕग़Ͱ͖ΔγεςϜΛߏங͢Δ w
ਖ਼൱ͷ༧ଌ͚ͩͰͳ͘طଘͷσʔλ͔Βநग़Ͱ͖Δ ใΛ༧ଌ͢Δ͜ͱΛࢼΈΔ !3
"VYJMJBSZ-PTT'VODUJPOT !4 xt h( f ) t h(b) t dt
yt DPSSFDUJODPSSFDU 8PSE yt,k : MBCFMLΛ࣋ͭτʔΫϯUͷ༧ଌ֬ ˜ yt,k : ̍τʔΫϯUͷਖ਼͍͠ϥϕϧ͕Lͷ࣌ ̌ͦΕҎ֎ͷ߹
"VYJMJBSZ-PTT'VODUJPOT !5 xt y(1) t y(2) t d(1) t d(2)
t h( f ) t h(b) t DPSSFDUJODPSSFDU 8PSE ผͷλεΫ yt,k : MBCFMLΛ࣋ͭτʔΫϯUͷ༧ଌ֬ ˜ yt,k : ̍τʔΫϯUͷਖ਼͍͠ϥϕϧ͕Lͷ࣌ ̌ͦΕҎ֎ͷ߹
"VYJMJBSZ-PTT'VODUJPOT w GSFRVFODZ ୯ޠͷසΛ༧ଌ͢Δ w FSSPSUZQF ୯ޠͷޡΓͷछྨΛ༧ଌ͢Δ w pSTUMBOHVBHF ֶशऀͷୈҰݴޠΛ༧ଌ͢Δ
w QBSUPGTQFFDI ୯ޠͷࢺΛ༧ଌ͢Δ w HSBNNBUJDBMSFMBUJPOT ୯ޠؒͷґଘؔΛ༧ଌ͢Δ ɹ !6 ผͷλεΫ
"VYJMJBSZ-PTT'VODUJPOT !7
&WBMVBUJPOTFUVQBOEEBUBTFUT w #BTFMJOFɿ3FJBOE:BOOBLPVEBLJT ɾ#J-45.Λ༻͍ͯ4P5"Λୡͨ͠Ϟσϧ w %BUBTFUT ɾ'JSTU$FSUJpDBUFJO&OHMJTI '$&
EBUBTFU ɾ$P/--TIBSFEUBTLUFTUTFU !8
&WBMVBUJPOTFUVQBOEEBUBTFUT w ύϥϝʔλ ɾXPSEFNCFEEJOHTTJ[F ɾJOJUJBMJ[FE QVCMJDMZBWBJMBCMFXPSEWFD .JLPMPWFUBM ɹFNCFEEJOHTUSBJOFEPO(PPHMF/FXT
ɾ-45.IJEEFOMBZFST ɾUBTLTQFDJpDIJEEFOMBZFST ɾPQUJNJ[FE"EBEFMUB ;FJMFS !9
!10
"MUFSOBUJWF5SBJOJOH4USBUFHJFT w ϚϧνλεΫֶशʹؔ͢Δݚڀ ෳͷσʔληοτͰಉ͡γεςϜΛ࠷దԽ͢Δ͜ͱʹ ॏΛஔ͍͍ͯΔ w ༗༻ੑΛࣔͨ͢Ίɺ࣍ͷσʔληοτΛ༻͍ͯ܇࿅ ɾ$P/--EBUBTFUʢDIVOLJOHʣ ɹɹ 5KPOH,JN4BOHBOE#VDIIPM[
ɾ$P/--DPSQVTʢ/&3ʣ ɹɹ 5KPOH,JN4BOHBOE%F.FVMEFS ɾ1FOO5SFFCBOL 15# 104DPSQVT ɹɹ .BSDVTFUBM !11
"MUFSOBUJWF5SBJOJOH4USBUFHJFT w ࣍ͷͭͷํ๏Ͱ܇࿅͢Δ ผͷλεΫͷσʔληοτͰ܇࿅ͨ͋͠ͱɺ ޡΓݕग़σʔληοτͰ܇࿅ ผͷλεΫͷσʔληοτͱޡΓݕग़ͷ σʔληοτΛަޓʹ܇࿅ !12
"MUFSOBUJWF5SBJOJOH4USBUFHJFT !13
"MUFSOBUJWF5SBJOJOH4USBUFHJFT !14
"EEJUJPOBM5SBJOJOH%BUB w NVMUJUBTLMFBSOJOH ར༻ՄೳͳλεΫݻ༗ͷ܇࿅σʔλ͕গͳ͍࣌ʹޮՌ͕ ظ͞ΕΔ w େنͳσʔληοτΛ༻͍ͨ߹ͷޮՌΛݕূ !15
"EEJUJPOBM5SBJOJOH%BUB w ࣍ͷσʔληοτΛ༻ʢ߹ܭ.UPLFOTʣ ɾ$BNCSJEHF-FBSOFS$PSQVT $-$ /JDIPMMT ɾ/64$PSQVTPG-FBSOFS&OHMJTI /6$-& %BIMNFJFSFUBM
ɾ-BOHDPSQVT .J[VNPUPFUBM w܇࿅࣌ʹֶश͢ΔλεΫ ɾ&SSPS%FUFDU ɾ104UBHHJOH !16
"EEJUJPOBM5SBJOJOH%BUB !17
$PODMVTJPO w ݴޠֶशऀͷ࡞ͨ͠จͷޡΓݕΛվળ͢ΔͨΊʹ χϡʔϥϧܥྻϥϕϦϯάʹิॿతͳଛࣦؔΛ౷߹ w 104λάɺจ๏తؔɺޡΓͷछྨ͕ޡΓݕग़ʹ༗༻Ͱ Έ߹ΘͤΔ͜ͱͰ݁Ռ͕վળ w ར༻Մೳͳ܇࿅σʔλ͕ݶΒΕ͍ͯΔ͚࣌ͩͰͳ͘ ଟ͘ͷ܇࿅σʔλΛ༻ͨ͠߹Ͱ༗ޮ
!18