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
Simple_Unsupervised_Summarization_by_Contextual...
Search
Sponsored
·
Your Podcast. Everywhere. Effortlessly.
Share. Educate. Inspire. Entertain. You do you. We'll handle the rest.
→
MARUYAMA
December 08, 2019
0
190
Simple_Unsupervised_Summarization_by_Contextual_Matching.pdf
MARUYAMA
December 08, 2019
Tweet
Share
More Decks by MARUYAMA
See All by MARUYAMA
vampire.pdf
tmaru0204
0
190
Misspelling_Oblivious_Word_Embedding.pdf
tmaru0204
0
200
Controlling_Text_Complexity_in_Neural_Machine_Translation.pdf
tmaru0204
0
180
20191028_literature-review.pdf
tmaru0204
0
160
Hint-Based_Training_for_Non-Autoregressive_Machine_Translation.pdf
tmaru0204
0
150
Soft_Contextual_Data_Augmentation_for_Neural_Machine_Translation_.pdf
tmaru0204
0
180
An_Embarrassingly_Simple_Approach_for_Transfer_Learning_from_Pretrained_Language_Models_.pdf
tmaru0204
0
160
Addressing_Trobulesome_Words_in_Neural_Machine_Translation.pdf
tmaru0204
0
170
Simple_Unsupervised_Keyphrase_Extraction_using_Sentence_Embeddings.pdf
tmaru0204
0
200
Featured
See All Featured
Product Roadmaps are Hard
iamctodd
PRO
55
12k
How to audit for AI Accessibility on your Front & Back End
davetheseo
0
180
Unlocking the hidden potential of vector embeddings in international SEO
frankvandijk
0
170
Navigating the Design Leadership Dip - Product Design Week Design Leaders+ Conference 2024
apolaine
0
180
Sharpening the Axe: The Primacy of Toolmaking
bcantrill
46
2.7k
Learning to Love Humans: Emotional Interface Design
aarron
275
41k
Building Experiences: Design Systems, User Experience, and Full Site Editing
marktimemedia
0
410
WCS-LA-2024
lcolladotor
0
450
Why Our Code Smells
bkeepers
PRO
340
58k
The Psychology of Web Performance [Beyond Tellerrand 2023]
tammyeverts
49
3.3k
Improving Core Web Vitals using Speculation Rules API
sergeychernyshev
21
1.4k
How to Create Impact in a Changing Tech Landscape [PerfNow 2023]
tammyeverts
55
3.2k
Transcript
4JNQMF6OTVQFSWJTFE4VNNBSJ[BUJPO CZ$POUFYUVBM.BUDIJOH จݙհ
1BQFS ɾ"$- ɾIUUQTXXXBDMXFCPSHBOUIPMPHZ1 ɾQBHFTr
"CTUSBDU ⾣ͭͷݴޠϞσϧͷΈͷγϯϓϧͳڭࢣͳ͠ੜܕཁ ⾣ੜܕཁɾநग़ܕཁͷํͰ༗༻ੑΛࣔͨ͠ ɾ$POUFYUVBMNBUDIJOHNPEFM ɾ%PNBJOqVFODZNPEFM
*OUSPEVDUJPO ⾣ڭࢣͳ͠ཁ ⾣ෳࡶͳϞσϧɾֶशΛඞཁͱ͠ͳ͍ϞσϧΛఏҊ MFOHUIDPOUSPMMFEWBSJBUJPOBMBVUPFODPEFS HFOFSBUJWFBEWFSTBSJBMOFUXPSL ʜ FH
.PEFM ⾣ཁจɺ࣍ͷͭͷಛੑΛຬ͍ͨͯ͠Δඞཁ͕͋Δ ਖ਼֬ੑ 'BJUIGVMOFTT ݪจͷҙຯͱಉ༷ͷҙຯΛ࣋ͭ ྲྀெੑ 'MVFODZ จ๏తʹਖ਼͘͠ཧղͰ͖Δ P(y|x) ∝
pcm (y|x)pfm (y|x)λ ೖྗจॻ x ཁจ y ਖ਼֬ੑ pcm (y|x) ྲྀெੑ pfm (y|x)
$POUFYUVBM.BUDIJOH.PEFM ⾣ਖ਼֬ੑೖྗςΩετͱग़ྗ୯ޠͷίαΠϯྨࣅ pcm (y|x) = N ∏ n=1 qcm (yn
|y<n , x) ͱͨ͠ͱ͖ sω = maxj≥1 Sim(x1:j , ω) ͜͜Ͱ ग़ྗީิޠͱೖྗςΩετͱͷྨࣅΛ ω x1:j qcm (y1 = ω|x) = softmax(s)
$POUFYUVBM.BUDIJOH.PEFM ⾣ॲཧखॱ sω = maxj≥zn−1 Sim(x1:j , ω) ࣍ࣜΑΓ ྨࣅΛܭࢉ
zn−1 ʹରԠ͘ҐஔҎ͔߱͠ߟྀ͠ͳ͍ zn−1 yn−1 ୯ௐੑͷԾఆ ग़ྗ ୯ޠҐஔΛܭࢉ qcm zn ͕ೖྗจॻඌͱͳΔ·Ͱ܁Γฦ͠ zn
%PNBJO'MVFODZ.PEFM ⾣ྲྀெੑݴޠϞσϧ֬ ೖྗจॻʹదͨ͠ޠΛબ͢ΔΑ͏ग़ྗޠኮΛ੍ ݴޠϞσϧͷग़ྗޠኮ7ΛϘϩϊΠׂʹΑΓ੍͖ޠኮ$ʹϚοϐϯά ͋Δڑ্ۭؒͷҙͷҐஔʹஔ͞Εͨෳݸͷʢʣʹରͯ͠ɺ ಉҰڑ্ۭؒͷଞͷ͕Ͳͷʹ͍͔ۙʹΑͬͯྖҬ͚͞Εͨਤͷ͜ͱɻ IUUQTKBXJLJQFEJBPSHXJLJϘϩϊΠਤ 8JLJQFEJBϘϩϊΠਤ
%PNBJO'MVFODZ.PEFM ⾣ྲྀெੑݴޠϞσϧ֬ pfm (y|x) = N ∏ n=1 ∑ ω′∈N(yn
) lm(ω′|y<n ) ϘϊϩΠׂͷʹ ೖྗจॻ୯ޠΛ༻͍Δ ͷϘϩϊΠྖҬΛͱͨ͠ͱ͖ ݴޠϞσϧ֬ yn N(yn )
.PEFM ⾣ཁจɺ࣍ͷͭͷಛੑΛຬ͍ͨͯ͠Δඞཁ͕͋Δ ਖ਼֬ੑ 'BJUIGVMOFTT ݪจͷҙຯͱಉ༷ͷҙຯΛ࣋ͭ ྲྀெੑ 'MVFODZ จ๏తʹਖ਼͘͠ཧղͰ͖Δ P(y|x) ∝
pcm (y|x)pfm (y|x)λ ೖྗจॻ x ཁจ y ਖ਼֬ੑ pcm (y|x) ྲྀெੑ pfm (y|x)
&YQFSJNFOUBMTFUVQ ⾣.PEFM ⾣%BUBTFU ɾੜܕཁ&OHMJTI(JHBXPSEEBUB 3VTI ɾநग़ܕཁ(PPHMFEBUBTFU 'JMJQQPWB
ɾGPSXBSEMBOHVBHFNPEFMPG&-.P ɾMBZFST-45.NPEFM pcm (y|x) pfm (y|x)
2VBOUJUBUJWF3FTVMUT 5BCMFੜܕཁͷ݁Ռ 5BCMFநग़ܕཁͷ݁Ռ
"OBMZTJT ⾣ೖྗจॻͷ୯ޠநग़͚ͩͰͳ͘ ੜͰ͖͍ͯΔ
4VNNBSZ ⾣ͭͷݴޠϞσϧͷΈͷγϯϓϧͳڭࢣͳ͠ੜܕཁ ⾣ੜܕཁɾநग़ܕཁͷํͰ༗༻ੑΛࣔͨ͠ ɾ$POUFYUVBMNBUDIJOHNPEFM ɾ%PNBJOqVFODZNPEFM
None