Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Back-Translation_Sampling_by_Targeting_Difficult_Words_in_Neural_Machine_Translation.pdf

MARUYAMA
February 27, 2019
0
120

 Back-Translation_Sampling_by_Targeting_Difficult_Words_in_Neural_Machine_Translation.pdf

MARUYAMA

February 27, 2019
Tweet

More Decks by MARUYAMA

See All by MARUYAMA
vampire.pdf
tmaru0204
0
130
Misspelling_Oblivious_Word_Embedding.pdf
tmaru0204
0
130
Simple_Unsupervised_Summarization_by_Contextual_Matching.pdf
tmaru0204
0
130
Controlling_Text_Complexity_in_Neural_Machine_Translation.pdf
tmaru0204
0
120
20191028_literature-review.pdf
tmaru0204
0
120
Hint-Based_Training_for_Non-Autoregressive_Machine_Translation.pdf
tmaru0204
0
100
Soft_Contextual_Data_Augmentation_for_Neural_Machine_Translation_.pdf
tmaru0204
0
120
An_Embarrassingly_Simple_Approach_for_Transfer_Learning_from_Pretrained_Language_Models_.pdf
tmaru0204
0
110
Addressing_Trobulesome_Words_in_Neural_Machine_Translation.pdf
tmaru0204
0
110

Featured

See All Featured
Responsive Adventures: Dirty Tricks From The Dark Corners of Front-End
smashingmag
245
20k
Faster Mobile Websites
deanohume
300
30k
Art, The Web, and Tiny UX
lynnandtonic
290
19k
Embracing the Ebb and Flow
colly
80
4.2k
Building Adaptive Systems
keathley
32
1.9k
Scaling GitHub
holman
457
140k
Practical Orchestrator
shlominoach
183
9.7k
KATA
mclloyd
16
12k
The Mythical Team-Month
searls
216
42k
Building a Scalable Design System with Sketch
lauravandoore
457
32k
Docker and Python
trallard
35
2.7k
Writing Fast Ruby
sferik
622
60k

Transcript

  1. #BDL5SBOTMBUJPO4BNQMJOHCZ5BSHFUJOH%J⒏DVMU 8PSETJO/FVSBM.BDIJOF5SBOTMBUJPO .BS[JFI'BEBFFBOE$ISJTUPG.PO[ &./-1 QBHFTr -JUFSBUVSFSFWJFX /BHBPLB6OJWFSTJUZPG5FDIOPMPHZ5BLVNJ.BSVZBNB

  2. ֓ཁ ⾣຋༁ͷ࣭ͷ޲্ʹޮՌతͳٯ຋༁ͷαϯϓϦϯάख๏ͱ͸ʁ ⾣8.5OFXTUSBOTMBUJPOUBTLͰϥϯμϜαϯϓϦϯάͷ ٯ຋༁ʹൺ΂ɺ#-&6͕ %F&O  &O%F ޲্ ⾣໨తݴޠଆͰ༧ଌ͢Δͷ͕೉͍͠୯ޠΛத৺ʹɺ ٖࣅσʔλΛੜ੒͢Δख๏ΛఏҊ

  3. ٯ຋༁ʹΑΔσʔλ֦ு 4PVSDFTJEF 5BSHFUTJEF 1BSBMMFMDPSQVT 1BSBMMFMDPSQVT .POPMJOHVBM $PSQVT #BDLXBSE NPEFM 5SBJOJOH

  4. ٯ຋༁ʹΑΔσʔλ֦ு 4PVSDFTJEF 5BSHFUTJEF 1BSBMMFMDPSQVT 1BSBMMFMDPSQVT .POPMJOHVBM $PSQVT #BDLXBSE NPEFM

  5. ٯ຋༁ʹΑΔσʔλ֦ு 4PVSDFTJEF 5BSHFUTJEF 1BSBMMFMDPSQVT 1BSBMMFMDPSQVT .POPMJOHVBM $PSQVT #BDLXBSE NPEFM 4ZOUIFUJDEBUB

    (FOFSBUJPO

  6. ٯ຋༁ʹΑΔσʔλ֦ு 4ZOUIFUJDEBUB 4PVSDFTJEF 5BSHFUTJEF 1BSBMMFMDPSQVT 1BSBMMFMDPSQVT .POPMJOHVBM $PSQVT 'PSXBSE NPEFM

    5SBJOJOH

  7. ٯ຋༁ʹΑΔσʔλ֦ு 4ZOUIFUJDEBUB 5FTUEBUB 4PVSDFTJEF 5BSHFUTJEF 1BSBMMFMDPSQVT 1BSBMMFMDPSQVT .POPMJOHVBM $PSQVT 'PSXBSE

    NPEFM

  8. ٯ຋༁ʹΑΔσʔλ֦ு 4ZOUIFUJDEBUB 5FTUEBUB 4PVSDFTJEF 5BSHFUTJEF 1BSBMMFMDPSQVT 1BSBMMFMDPSQVT .POPMJOHVBM $PSQVT 3FTVMU

    'PSXBSE NPEFM

  9. ͸͡Ίʹ ⾣Ұൠతʹɺʮٯ຋༁ʹΑΔσʔλ֦ு͸ɺPWFSpUUJOH΍ ྲྀெੑͷ໰୊Λܰݮ͢ΔʯͱݴΘΕ͍ͯΔ ⾣ϙΠϯτ͸ɺʮͲͷΑ͏ͳσʔλΛ༻͍֦ͯுΛߦ͏͔ʯ

  10. ⾣ςετσʔλͱυϝΠϯ͕ྨࣅ͢ΔσʔλΛબ୒ͭͭ͠ ϞσϧΛֶशͤ͞Δ͜ͱͰੑೳΛվળ ؔ࿈ݚڀ ⾣ٯ຋༁ʹ͓͚ΔɺυϝΠϯʹج͍ͮͨσʔλબ୒͸ɺ ϥϯμϜαϯϓϦϯάͱൺ΂ͯେࠩͳ͠ WBOEFS8FFTFUBM 1IBNFUBM

  11. ⾣ٖࣅσʔλ௥Ճલޙʹ͓͚Δɺ༧ଌଛࣦͷมԽΛௐࠪ ٯ຋༁ͱτʔΫϯ༧ଌ "UUFOUJPOCBTFEFODPEFSEFDPEFSNPEFM ϥϯμϜαϯϓϦϯάʹΑΔٯ຋༁Ͱٙࣅσʔλੜ੒

  12. ٯ຋༁ͱτʔΫϯ༧ଌ ٖࣅσʔλ௥ՃʹΑΔ ༧ଌଛࣦͷ૿ݮ ໨తݴޠଆͷ୯ޠස౓ ର਺

  13. ٯ຋༁ͱτʔΫϯ༧ଌ ٖࣅσʔλ௥ՃʹΑΔ ༧ଌଛࣦͷݮগ͸ CBTFMJOFQSFEJDUJPOMPTT ͕ߴ͍୯ޠͰൃੜ͢Δ ʮ༧ଌ͕ࠔ೉ʯͳ΋ͷΛ ٙࣅσʔλʹՃ͑ͨํ͕ ޮՌతʁ

  14. ༧ଌࠔ೉ͳτʔΫϯʹجͮ͘αϯϓϦϯά ⾣ʮ༧ଌ͕ࠔ೉ʯͳ΋ͷΛͲͷΑ͏ʹαϯϓϦϯά͢Δ͔ʁ ස౓ͷᮢ஋ ༧ଌଛࣦͷฏۉͷᮢ஋ ༧ଌଛࣦͷඪ४ภࠩͷᮢ஋ 1SFEJDUJPOMPTTͷൺ཰ʹ߹ΘͤͯαϯϓϦϯά

  15. ༧ଌࠔ೉ͳτʔΫϯʹجͮ͘αϯϓϦϯά

  16. จ຺ʹجͮ͘αϯϓϦϯά ⾣ಉ͡τʔΫϯͰ΋ʮ༧ଌͷࠔ೉͞ʯ͸จ຺ʹΑͬͯҟͳΔ ⾣ྨࣅͨ͠จ຺Λ΋ͭ৔߹͸ɺͦͷτʔΫϯͷ༧ଌͷࠔ೉͞ ʹ͍ͭͯ΋ྨࣅ͍ͯ͠ΔͱԾఆ ⾣ʮ༧ଌͷࠔ೉͞ʯ ʮจ຺ͷྨࣅੑʯͰαϯϓϦϯά

  17. จ຺ʹجͮ͘αϯϓϦϯά ⾣lจ຺zͷఆٛ /FJHICPSJOHUPLFOT 50,&/4 /FJHICPSJOHTVCXPSEVOJUT 4803%4 

  18. จ຺ʹجͮ͘αϯϓϦϯά ⾣lྨࣅ౓zͷܭࢉํ๏ .BUDIJOHUIFMPDBMDPOUFYU ."5$) 8PSESFQSFTFOUBUJPOT &.#  7 ɾ จ຺಺ͷ֤τʔΫϯͷ୯ޠϕΫτϧͷฏۉ

  19. จ຺ʹجͮ͘αϯϓϦϯά

  20. ·ͱΊ ⾣຋༁ͷ࣭ͷ޲্ʹޮՌతͳٯ຋༁ͷαϯϓϦϯάख๏ͱ͸ʁ ⾣໨తݴޠଆͰ༧ଌ͢Δͷ͕೉͍͠୯ޠΛத৺ʹɺ ٖࣅσʔλΛੜ੒͢Δख๏ΛఏҊ ⾣8.5OFXTUSBOTMBUJPOUBTLͰϥϯμϜαϯϓϦϯάͷ ٯ຋༁ʹൺ΂ɺ#-&6͕ %F&O  &O%F ޲্

  21. ࢀߟจݙ <>/HPD2VBO1IBN +BO/JFIVFT 5IBOI-F)B &VOBI$IP .BUUIJBT4QFSCFS BOE "MFYBOEFS8BJCFM5IFLBSMTSVIFJOTUJUVUFPGUFDIOPMPHZTZTUFNTGPSUIFOFXT USBOTMBUJPOUBTLJOXNU*O1SPDFFEJOHTPGUIF4FDPOE$POGFSFODFPO.BDIJOF 5SBOTMBUJPO

    QBHFTr <>.BSMJFTWBOEFS8FFT "SJBOOB#JTB[[B BOE$ISJTUPG.PO[%ZOBNJDEBUB TFMFDUJPOGPSOFVSBMNBDIJOFUSBOTMBUJPO*O1SPDFFEJOHTPGUIF$POGFSFODFPO &NQJSJDBM.FUIPETJO/BUVSBM-BOHVBHF1SPDFTTJOH QBHFTr $PQFOIBHFO  %FONBSL"TTPDJBUJPOGPS$PNQVUBUJPOBM-JOHVJTUJDT