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
Addressing_Trobulesome_Words_in_Neural_Machine_...
Search
MARUYAMA
May 28, 2019
180
0
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
Addressing_Trobulesome_Words_in_Neural_Machine_Translation.pdf
MARUYAMA
May 28, 2019
More Decks by MARUYAMA
See All by MARUYAMA
vampire.pdf
tmaru0204
0
200
Misspelling_Oblivious_Word_Embedding.pdf
tmaru0204
0
210
Simple_Unsupervised_Summarization_by_Contextual_Matching.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
Simple_Unsupervised_Keyphrase_Extraction_using_Sentence_Embeddings.pdf
tmaru0204
0
200
Featured
See All Featured
Applied NLP in the Age of Generative AI
inesmontani
PRO
4
2.3k
CoffeeScript is Beautiful & I Never Want to Write Plain JavaScript Again
sstephenson
162
16k
Fight the Zombie Pattern Library - RWD Summit 2016
marcelosomers
234
17k
Intergalactic Javascript Robots from Outer Space
tanoku
273
27k
Hiding What from Whom? A Critical Review of the History of Programming languages for Music
tomoyanonymous
2
870
Being A Developer After 40
akosma
91
590k
Build your cross-platform service in a week with App Engine
jlugia
234
18k
Visualization
eitanlees
152
17k
Leading Effective Engineering Teams in the AI Era
addyosmani
9
2.1k
A Soul's Torment
seathinner
6
3k
The Hidden Cost of Media on the Web [PixelPalooza 2025]
tammyeverts
2
340
Effective software design: The role of men in debugging patriarchy in IT @ Voxxed Days AMS
baasie
0
440
Transcript
"EESFTTJOH5SPVCMFTPNF8PSET JO/FVSBM.BDIJOF5SBOTMBUJPO :BOH;IBP +JBKVO;IBOH ;IPOHKVO)F $IFOHRJOH;POH BOE)VB8V &./-1 QBHFT -JUFSBUVSFSFWJFX
/BHBPLB6OJWFSTJUZPG5FDIOPMPHZ5BLVNJ.BSVZBNB
"CTUSBDU ⾣$O&O %F&Oʹ͓͍ͯɺ༁࣭ͷվળΛ֬ೝ ⾣ॲཧͷ͍͠୯ޠ 5SPVCMFTPNFXPSE ͷఆٛ ⾣χϡʔϥϧػց༁ /.5 ʹ͓͍ͯɺॲཧͷ͍͠୯ޠΛదʹ ɹऔΓѻ͓͏ͱ͍͏ࢼΈ
⾣5SPVCMFTPNFXPSEͱͦͷจ຺Λߟྀͨ͠ϝϞϦػߏ
5SPVCMFTPNF8PSE%FpOJUJPO ⾣ॲཧͷखॱ ༧Ίֶश͓͍ͯͨ͠/.5Λ༻͍ͯɺ ɹHPMEUBSHFUXPSEͷ֬Λࢉग़ ୯ޠΞϥΠϝϯτπʔϧΛ༻͍ͯɺ ɹHPMEUBSHFUXPSEʹରԠͮ͘ɺ TPVSDFXPSEΛٻΊΔ &YDFQUJPODSJUFSJPOΛຬͨͨ͠߹ɺ ɹͦͷTPVSDFXPSEΛ
USPVCMFTPNFXPSEͱΈͳ͢ ᶃ ᶄ ᶅ
5SPVCMFTPNF8PSE%FpOJUJPO ⾣&YDFQUJPO$SJUFSJB "CTPMVUF$SJUFSJPO (BQ$SJUFSJPO 3BOLJOH$SJUFSJPO 1SPCBCJMJUZ EJTUSJCVUJPO Source sentence =
(x1 , x2 , x3 ) Target sentence = (y1 , y2 , y3 ) HPMETUBOEBSE
5SPVCMFTPNF8PSE%FpOJUJPO ⾣&YDFQUJPO$SJUFSJB "CTPMVUF$SJUFSJPO HPMETUBOEBSEͷग़ྗ֬ͷᮢ PN i (yi ) < p0
1SPCBCJMJUZ EJTUSJCVUJPO Source sentence = (x1 , x2 , x3 ) Target sentence = (y1 , y2 , y3 ) HPMETUBOEBSE PN 1 (y1 ) = 0.80 PN 2 (y2 ) = 0.31 PN 3 (y3 ) = 0.20 Y Y ͕USPVCMFTPNFXPSE ͳΒɺ p0 = 0.50 FH
5SPVCMFTPNF8PSE%FpOJUJPO ⾣&YDFQUJPO$SJUFSJB ͱͷࠩͷᮢ gi (yi ) = max(PN i (yi
)) − PN i (yi ) 1SPCBCJMJUZ EJTUSJCVUJPO Source sentence = (x1 , x2 , x3 ) Target sentence = (y1 , y2 , y3 ) HPMETUBOEBSE g1 (y1 ) = 0.80 − 0.80 = 0.00 g2 (y2 ) = 0.35 − 0.31 = 0.04 g3 (y3 ) = 0.75 − 0.20 = 0.55 Y ͕USPVCMFTPNFXPSE ͳΒɺ g0 = 0.10 FH max(PN i (yi )) PN i (yi ) gi (yi ) > g0 (BQ$SJUFSJPO
5SPVCMFTPNF8PSE%FpOJUJPO ⾣&YDFQUJPO$SJUFSJB ग़ྗ֬ॱҐͷᮢ 1SPCBCJMJUZ EJTUSJCVUJPO Source sentence = (x1 ,
x2 , x3 ) Target sentence = (y1 , y2 , y3 ) HPMETUBOEBSE rank(y1 ) = 1 rank(y2 ) = 3 rank(y3 ) = 2 Y ͕USPVCMFTPNFXPSE ͳΒɺ rank0 = 2 FH rank(yi ) > rank0 3BOLJOH$SJUFSJPO
*OUFHSBUJOH$POUFYUVBM.FNPSZJOUP/.5 ⾣༁ॲཧͷྲྀΕ NFNPSZΑΓUSPVCMFTPNFXPSEΛநग़ DPOUFYUVBMTJNJMBSJUZΛܭࢉ NFNPSZQSFEJDUFEQSPCBCJMJUZͷܭࢉ ग़ྗ֬Λܭࢉ ᶃ ᶄ ᶅ
ᶆ
*OUFHSBUJOH$POUFYUVBM.FNPSZJOUP/.5 ⾣༁ॲཧͷྲྀΕ NFNPSZΑΓUSPVCMFTPNFXPSEΛநग़ ᶃ sm USPVCMFTPNFTPVSDFXPSE tn HPMEUBSHFUXPSE c(sm
, tn ) IJEEFOTUBUFPGFODPEFS pL(sm , tn ) MFYJDPOUSBOTMBUJPOQSPCBCJMJUZ
*OUFHSBUJOH$POUFYUVBM.FNPSZJOUP/.5 ⾣༁ॲཧͷྲྀΕ DPOUFYUVBMTJNJMBSJUZΛܭࢉ NFNPSZQSFEJDUFEQSPCBCJMJUZͷܭࢉ ग़ྗ֬Λܭࢉ ᶅ ᶆ ᶄ
&YQFSJNFOUBM4FUUJOHT ⾣%BUBTFUT 5SBJOJOH-%$DPSQVT .TFOUFODFQBJST 7BMJEBUJPO/*45 5FTU/*45 ⾣.PEFMT #BTFMJOF/.5XJUIHMPCBMBUUFOUJPO "SUIVS/.5
MFYJDBMUSBOTMBUJPO 9 .&.ఏҊख๏
3FTVMUT ⾣ϝϞϦػߏͷޮՌ
3FTVMUT ⾣&YDFQUJPODSJUFSJBʹΑΔҧ͍
3FTVMUT ⾣සޠ ᐆດੑͷ͋ΔޠΛؚΉจͰͷධՁ -PXग़ݱճ͕ճҎԼͷ୯ޠ "NCΤϯτϩϐʔ͕ᮢҎ্ − K ∑ k=1 pL
k logpL k > E0 (E0 = 1.5)
3FTVMUT ⾣සޠ ᐆດੑͷ͋ΔޠΛؚΉจͰͷධՁ #-&6είΞͷൺֱ
3FTVMUT ⾣සޠ ᐆດੑͷ͋ΔޠΛؚΉจͰͷධՁ 5XPSEUSPVCMFTPNFXPSE &SSPS#BTFMJOFʹ͓͚Δؒҧͬͨग़ྗ 3FDUJGZఏҊख๏ಋೖʹΑΓमਖ਼Ͱ͖ͨ %FUFSJPఏҊख๏ಋೖʹΑΓ૿͑ͨؒҧ͍ͷ
$PODMVTJPO ⾣$O&O %F&Oʹ͓͍ͯɺ༁࣭ͷվળΛ֬ೝ ⾣ॲཧͷ͍͠୯ޠ 5SPVCMFTPNFXPSE ͷఆٛ ⾣χϡʔϥϧػց༁ /.5 ʹ͓͍ͯɺॲཧͷ͍͠୯ޠΛదʹ ɹऔΓѻ͓͏ͱ͍͏ࢼΈ
⾣5SPVCMFTPNFXPSEͱͦͷจ຺Λߟྀͨ͠ϝϞϦػߏ