$30 off During Our Annual Pro Sale. View Details »
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
Dynamic Multi-Level Multi-Task Learning for Sen...
Search
onizuka laboratory
October 17, 2018
Research
0
54
Dynamic Multi-Level Multi-Task Learning for Sentence Simplification
弊研究室で行なったCOLING2018読み会の発表資料です。
onizuka laboratory
October 17, 2018
Tweet
Share
More Decks by onizuka laboratory
See All by onizuka laboratory
Phrase-Based & Neural Unsupervised Machine Translation
onilab
0
120
Tell-and-Answer: Towards Explainable Visual Question Answering using Attributes and Captions
onilab
0
72
Card-660: A Reliable Evaluation Framework for Rare Word Representation Models
onilab
0
36
A Word-Complexity Lexicon and A Neural Readability Ranking Model for Lexical Simplification
onilab
0
130
Integrating Transformer and Paraphrase Rules for Sentence Simplification
onilab
0
61
An Auto-Encoder Matching Model for Learning Utterance-Level Semantic Dependency in Dialogue Generation
onilab
0
57
Generating More Interesting Responses in Neural Conversation Models with Distributional Constraints
onilab
0
100
Modeling Multi-turn Conversation with Deep Utterance Aggregation
onilab
0
98
Learning Semantic Sentence Embeddings using Pair-wise Discriminator
onilab
0
120
Other Decks in Research
See All in Research
空間音響処理における物理法則に基づく機械学習
skoyamalab
0
110
思いつきが武器になる:研究というゲームを始めよう / Ideas Are Your Equipments : Let the Game of Research Begin!
ks91
PRO
0
100
多言語カスタマーインタビューの“壁”を越える~PMと生成AIの共創~ 株式会社ジグザグ 松野 亘
watarumatsuno
0
170
AIスパコン「さくらONE」の オブザーバビリティ / Observability for AI Supercomputer SAKURAONE
yuukit
2
990
大規模言語モデルにおけるData-Centric AIと合成データの活用 / Data-Centric AI and Synthetic Data in Large Language Models
tsurubee
1
440
その推薦システムの評価指標、ユーザーの感覚とズレてるかも
kuri8ive
1
270
snlp2025_prevent_llm_spikes
takase
0
420
湯村研究室の紹介2025 / yumulab2025
yumulab
0
240
POI: Proof of Identity
katsyoshi
0
120
Language Models Are Implicitly Continuous
eumesy
PRO
0
350
能動適応的実験計画
masakat0
2
1.1k
EcoWikiRS: Learning Ecological Representation of Satellite Images from Weak Supervision with Species Observation and Wikipedia
satai
3
430
Featured
See All Featured
Building an army of robots
kneath
306
46k
Agile that works and the tools we love
rasmusluckow
331
21k
Principles of Awesome APIs and How to Build Them.
keavy
127
17k
Making the Leap to Tech Lead
cromwellryan
135
9.7k
Understanding Cognitive Biases in Performance Measurement
bluesmoon
32
2.7k
How STYLIGHT went responsive
nonsquared
100
6k
Fight the Zombie Pattern Library - RWD Summit 2016
marcelosomers
234
17k
Making Projects Easy
brettharned
120
6.5k
Fireside Chat
paigeccino
41
3.7k
Designing for humans not robots
tammielis
254
26k
The Art of Delivering Value - GDevCon NA Keynote
reverentgeek
16
1.8k
Save Time (by Creating Custom Rails Generators)
garrettdimon
PRO
32
1.8k
Transcript
$ )BO(VP 3BNBLBOUI 1BTVOVSV .PIJU #BOTBM %ZOBNJD.VMUJ-FWFM.VMUJ5BTL -FBSOJOHGPS4FOUFODF 4JNQMJGJDBUJPO
#ݪ େو $0-*/(ಡΈձʢʣ
࣍ l ֓ཁ l Ϟσϧ l ධՁํ๏ l ݁Ռ l
"CMBUJPOT"OBMZTJT
֓ཁ จͷฏқԽ l ׂɺআɺݴ͍͑ͳͲͰɺՄಡੑΛ্͛Δ l ༗ޮͳೖྗจཧతʹؚΊΔ͖ʢؚҙʣ #BTFMJOF l ϙΠϯλίϐʔϝΧχζϜͷTFRTFR
֓ཁ ఏҊख๏ l ϚϧνλεΫɾϚϧνϨϕϧ֊ ◦ ิॿλεΫʢؚҙɾݴ͍͑ʣ͕ɺҟͳΔϨϕϧͷ֊ l λεΫͷΓସ͑ํΛಈతʹֶश ◦ ଟόϯσΟοτϕʔεͷ܇࿅ख๏
ධՁɾੳ l ࣗಈධՁʢ4"3*ɺ',(-ʣͱखಈධՁͰ༏Εͨ
࣍ l ֓ཁ l Ϟσϧ l ධՁํ๏ l ݁Ռ l
"CMBUJPOT"OBMZTJT Ø Baseline Ø Ø Ø Ø Ø
Ϟσϧɿ#BTFMJOF ϙΠϯλίϐʔจฏқԽϞσϧ l ҎԼͷΛඋ͑ͨTFRTFRϞσϧ ◦ Ξςϯγϣϯʢ#BIEBOBV FUBM ʣ ◦ ϙΠϯλίϐʔϝΧχζϜʢ4FFFUBM
ʣ
Ϟσϧɿ#BTFMJOF Ξςϯγϣϯ 4FFFUBM l(FU5P5IF1PJOU4VNNBSJ[BUJPOXJUI1PJOUFS(FOFSBUPS/FUXPSLTz BS9JWF
Ϟσϧɿ#BTFMJOF ϙΠϯλίϐʔϝΧχζϜ 4FFFUBM l(FU5P5IF1PJOU4VNNBSJ[BUJPOXJUI1PJOUFS(FOFSBUPS/FUXPSLTz BS9JWF
ϞσϧɿิॿλεΫ̍ ؚҙੜλεΫ l ฏқจೖྗʹै͏͖ l ҙຯతʢߴʣ l ؚҙੜϞσϧ܇࿅༻σʔληοτ ◦ 4/-*
#PXNBOFUBM ◦ .VMUJ/-* 8JMMJBNTFUBM
ϞσϧɿิॿλεΫ̎ ݴ͍͑ੜλεΫ l ߏจޠኮͰɺฒସ͑ॻ͖͑ l ޠኮ౷ޠʢʣ l ݴ͍͑ੜλεΫ༻܇࿅σʔληοτ ◦ 1BSB/.5
8JFUJOH BOE(JNQFM B
ϞσϧɿϚϧνλεΫϞσϧ
ϞσϧɿϚϧνλεΫֶश TFRTFRͰ l Τϯίʔμσίʔμͷɺύϥϝʔλڞ༗ ఏҊख๏ l ڞ༗ͰɺύϥϝʔλͷҰ෦Λඇެ։ɺ Γʢؔ࿈දݱʣΛڞ༗ l ڞ༗ύϥϝʔλΛଛࣦؔͷϖφϧςΟ߲Ͱɺ
͋Δڑࢦඪʹ͚ۙͮΔ
ϞσϧɿϚϧνϨϕϧڞ༗ TFRTFRͷ֤֊ͰҟͳΔػೳΛͭ l #FMJOLPW FUBM l ʢೖྗଆʣʹ୯ޠߏΛֶशʢݴ͍͑ʣ l ߴʹҙຯʹযΛ߹ΘͤΔʢؚҙʣ
Ϟσϧɿιϑτڞ༗ ϋʔυڞ༗ l ڞ༗͢ΔύϥϝʔλΛ݁ͼ͚ͭΔ ιϑτڞ༗ l ҟͳΔλεΫ͕ɺύϥϝʔλۭؒͷͲͷ෦Λڞ༗͢Δ ͔બͰ͖Δ l ଛࣦؔ
! " = −log () * +; " + . "/ − 0/ ◦ " ओλεΫʢฏқԽʣͷશͯͷύϥϝʔλ ◦ "/ 0/ ओλεΫͱิॿλεΫͷڞ༗ύϥϝʔλͷαϒηοτ
ϞσϧɿϚϧνλεΫ܇࿅ طଘख๏ l ʢ੩తͳʣࠞ߹ൺ !"" : !$% : !&& ◦
ओλεΫʢฏқԽʣɿؚҙɿݴ͍͑ ఏҊख๏ l ੩తͰͳ͘ಈతʹ܇࿅͍ͨ͠
Ϟσϧɿಈతࠞ߹ൺֶश ଟόϯσΟοτ l XJUI#PMU[NBOOFYQMPSBUJPO ,BFMCMJOH FUBM l XJUIࢦҠಈฏۉߋ৽ϧʔϧ
Ϟσϧɿಈతࠞ߹ൺֶश ଟόϯσΟοτ
Ϟσϧɿಈతࠞ߹ൺֶश ଟόϯσΟοτ l ̏λεΫͷઃఆͷू߹ !" , ⋯ , !% l
֤ &' ͰɺΛબ͠ɺใु ()* ΛಘΔ ◦ ใुɺओλεΫͷෛͷ WBMJEBUJPOMPTT l όϯσΟοτίϯτϩʔϥͷํ l ֤ &' Ͱͷ֤ + ͷߦಈਪఆ
࣍ l ֓ཁ l Ϟσϧ l ධՁํ๏ l ݁Ռ l
"CMBUJPOT"OBMZTJT
ධՁํ๏ɿσʔληοτ 5SBJO 7BMJE 5FTU /FXTFMB 4NBMM8JLJ
-BSHF8JLJ 4/-* .VMUJ/-* ؚҙੜ༻ ʢ߹Θͤͯʣ 8JFUJOH(JNQFM ݴ͍͑༻ . ʢ߹Θͤͯʣ,
ධՁํ๏ɿධՁࢦඪ ࣗಈධՁ l 4"3* l ',(- ◦ ͍จΛධՁ͕ͪ͠ʢ4IBSEMPX ʣ l
#-&6 ◦ ฏқԽͰؔ࿈ੑ͕͍ʢ;IVFUBM ଞʣ ◦ มߋ͠ͳ͍อकతͳγεςϜ͕ධՁ͞Ε͕ͪʢಉ্ʣ
ධՁํ๏ɿධՁࢦඪ खಈධՁ l ྲྀெੑʢ'MVFODZʣ ◦ JTUIFPVUQVUHSBNNBUJDBMBOEXFMMGPSNFE l ଥੑʢ"EFRVBDZʣ ◦ UPXIBUFYUFOUJTUIFNFBOJOHFYQSFTTFEJOUIFPSJHJOBM
TFOUFODFQSFTFSWFEJOUIFPVUQVU l ฏқੑʢ4JNQMJDJUZʣ ◦ JTUIFPVUQVUTJNQMFSUIBOUIFPSJHJOBMTFOUFODF
࣍ l ֓ཁ l Ϟσϧ l ධՁํ๏ l ݁Ռ l
"CMBUJPOT"OBMZTJT
݁Ռ l #BTFMJOFWTطଘख๏ ◦ /FXTFMB Ͱ ',(-༏উɺ4"3*࣍ ◦ 8JLJͰಉ
݁Ռ l #BTFMJOF &OU1BSWT#BTFMJOFطଘख๏ ◦ /FXTFMB ͷ ',(-ͱ 4"3*Ͱ༏উ l
#BTFMJOF &OU 1BSWT#BTFMJOF &OU1BS ◦ /FXTFMB 8JLJͰ ',(- 4"3*͍͍ͩͨ༏উ
݁Ռ l ʢิॿλεΫͷʣࠞ߹ൺʢಈతWT੩తʣ ◦ /FXTFMB ͱ 8JLJ4NBMM Ͱ 4"3* ༏উ
݁Ռ ਓखධՁ l XBZWTଞ ◦ ฏқੑʢ4JNQMJDJUZʣͰ༏উ ◦ ଥੑʢ"EFRVBDZʣ %3&44-4 ʹෛ͚ͨˠฏқԽॏࢹ
◦ ฏۉείΞɺ%3&44-4ΑΓྑ͍
݁Ռ ೖग़ྗͷҰக l XBZ͍ʢ (SPVOEUSVUIʹ͍ۙʣ l %3&44-4 ߴ͍ ◦ ଥੑʢ"EFRVBDZʣ͕ߴ͔ͬͨͷᰐ͚Δ
࣍ l ֓ཁ l Ϟσϧ l ධՁํ๏ l ݁Ռ l
"CMBUJPOT"OBMZTJT
"CMBUJPOT"OBMZTJT ଞͷڞ༗ख๏ʁ l ʴߴ͕༏উ
"CMBUJPOT"OBMZTJT ڞ༗ʢιϑτ WTϋʔυʣʁ l ιϑτ͕ 4"3*ڧ͍
"CMBUJPOT"OBMZTJT 4"3*ͷৄࡉ l XBZ༏উ
"CMBUJPOT"OBMZTJT ̎ͭͷଟόϯσΟοτख๏ l ͭɿ࠷ޙΛهͯࠞ͠߹ൺΛݻఆ͢Δ l 4"3*ʢ̍ͭʣWT ʢ̎ͭʣ
"CMBUJPOT"OBMZTJT ̎ͭͷଟόϯσΟοτख๏
"CMBUJPOT"OBMZTJT ϚϧνλεΫֶश WTσʔλ֦ு l ຊʹɺิॿλεΫͷzػೳz͕ੑೳ্ͤͨ͞ʁ ◦ ิॿλεΫ༻ͷlσʔλՃzͷӨڹͰͳ͍ʁ l ิॿλεΫ༻σʔλʢ4/-* .VMUJ/-*
ͱ 1BSB/.5ʣΛɺݸผʹ FNCFEEJOHͯ͠ ओλεΫϞσϧʹΈࠐΜͩ l ఏҊख๏͕େ෯ʹ༏Ε͍ͯͨ
"CMBUJPOT"OBMZTJT ྫ