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

[論紹介] Fast Lexically Constrained Decoding with Dynamic Beam Allocation for Neural Machine Translation

onizuka laboratory
July 11, 2018
Research
0
340

[論紹介] Fast Lexically Constrained Decoding with Dynamic Beam Allocation for Neural Machine Translation

弊研究室で行なったNAACL読み会の発表資料です。

onizuka laboratory

July 11, 2018
Tweet

More Decks by onizuka laboratory

See All by onizuka laboratory
Phrase-Based & Neural Unsupervised Machine Translation
onilab
0
110
Tell-and-Answer: Towards Explainable Visual Question Answering using Attributes and Captions
onilab
0
65
Card-660: A Reliable Evaluation Framework for Rare Word Representation Models
onilab
0
31
A Word-Complexity Lexicon and A Neural Readability Ranking Model for Lexical Simplification
onilab
0
98
Integrating Transformer and Paraphrase Rules for Sentence Simplification
onilab
0
56
An Auto-Encoder Matching Model for Learning Utterance-Level Semantic Dependency in Dialogue Generation
onilab
0
48
Generating More Interesting Responses in Neural Conversation Models with Distributional Constraints
onilab
0
98
Modeling Multi-turn Conversation with Deep Utterance Aggregation
onilab
0
91
Learning Semantic Sentence Embeddings using Pair-wise Discriminator
onilab
0
110

Other Decks in Research

See All in Research
CSC590 Lecture 01
javiergs
PRO
0
130
SSII2023 医療支援における画像処理研究の動向と展望
moda0
0
110
プロシェアリング白書2024_PROSHARING_REPORT_2024
circulation
0
660
機械学習と最適化の融合 動的ロットサイズ決定問題を例として
mickey_kubo
1
220
Breaking Tradeoffs: Extremely Scalable Multi-Agent Pathfinding Algorithms
kei18
0
150
論文紹介 DSRNet: Single Image Reflection Separation via Component Synergy (ICCV 2023)
tattaka
0
180
言語間転移学習で大規模言語モデルを賢くする
ikuyamada
7
3.5k
Combating Misinformation in the age of LLMs
teacherpeterpan
0
140
LiDARセキュリティ最前線
kentaroy47
0
280
10-ot-generic-bio.pdf
gpeyre
0
140
方策の長期性能に対する 効率的なオフライン評価・学習 (Long-term Off-Policy Evaluation and Learning)
usaito
PRO
2
190
Cross-Media Information Spaces and Architectures
signer
PRO
0
120

Featured

See All Featured
Cheating the UX When There Is Nothing More to Optimize - PixelPioneers
stephaniewalter
275
13k
The Success of Rails: Ensuring Growth for the Next 100 Years
eileencodes
33
6k
Making the Leap to Tech Lead
cromwellryan
125
8.5k
XXLCSS - How to scale CSS and keep your sanity
sugarenia
242
1.2M
Embracing the Ebb and Flow
colly
80
4.2k
Build The Right Thing And Hit Your Dates
maggiecrowley
25
2k
Designing on Purpose - Digital PM Summit 2013
jponch
111
6.5k
It's Worth the Effort
3n
180
27k
Being A Developer After 40
akosma
66
580k
Writing Fast Ruby
sferik
622
60k
From Idea to $5000 a Month in 5 Months
shpigford
378
45k
Unsuck your backbone
ammeep
663
57k

Transcript

  1. 'BTU-FYJDBMMZ$POTUSBJOFE%FDPEJOH XJUI%ZOBNJD#FBN"MMPDBUJPO GPS/FVSBM.BDIJOF5SBOTMBUJPO .BUU1PTUBOE%BWJE7JMBS "NB[PO3FTFBSDI #FSMJO (FSNBOZ 1SPDPG/""$- QQ 1SFTFOUFECZ5PNPZVLJ,BKJXBSB

  2. $POUSPMMJOH/.5PVUQVU • /.5SFNPWFTNBOZXBZTPGNBOVBMMZHVJEJOH UIFUSBOTMBUJPOQSPDFTTJOPMEFSQBSBEJHNT 1#4.5  • *ODPSQPSBUFEPNBJOTQFDJGJDEJDUJPOBSJFT • 'PSDFBUSBOTMBUJPODIPJDFGPSDFSUBJOXPSET

    • -FYJDBMMZ$POTUSBJOFE%FDPEJOHGPS4FRVFODF(FOFSBUJPO 6TJOH(SJE#FBN4FBSDI )PLBNQ BOE-JV "$- • 6TFSTDBOTQFDJGZUIFXPSETUIBUNVTUBQQFBSJOUIFPVUQVU • 6TFGVMGPSUBTLTTVDIBTQPTUFEJUJOH DBQUJPOHFOFSBUJPO   /JFNBOE IBUEJF"CTJDIU FJOF .BVFS[V CBVFO /JFNBOE IBUEJF"CTJDIU FJOF .BVFS[V FSSJDIUFO ,FJOFS IBUEJF"CTJDIU FJOF .BVFS[V CBVFO FSSJDIUFO LFJOFS -FYJDBMMZ$POTUSBJOFE%FDPEJOH /PPOFIBTUIFJOUFOUJPO PGCVJMEJOHBXBMM

  3. (#4IBTBMBSHF3VOUJNF$PNQMFYJUZ %FDPEJOH$PNQMFYJUZ 0 /L$ (SJE#FBN4FBSDI )PLBNQ BOE-JV  0 /L

    %ZOBNJD#FBN"MMPDBUJPO 5IJT8PSL /TFOUFODFMFOHUI LCFBNTJ[F $OVNCFSPGDPOTUSBJOUT (#4IBTUXPQSPCMFNT • %FDPEJOHDPNQMFYJUZJTMJOFBSJODPOTUSBJOUT • $BOOPUTQFDJGZUIFCFBNTJ[FBUNPEFMMPBEUJNF 

  4. %#"%ZOBNJD#FBN"MMPDBUJPO • 4JNQMZSFQMBDFUIF,#&45 JNQMFNFOUBUJPO GSPNUIFBMHPSJUINPGTUBOEBSECFBNEFDPEJOH • *OTUFBEPGTFMFDUJOHUIFUPQLJUFNT  (FOFSBUJOHBMJTUPGDBOEJEBUFT 8FOPXOFFEUPFOTVSFUIBUDBOEJEBUFTQSPHSFTT

    UISPVHIUIFTFUPGQSPWJEFEDPOTUSBJOUT  "MMPDBUJOHUIFCFBNBDSPTTUIFDPOTUSBJOUCBOLT 8JUIBGJYFETJ[FECFBNBOEBSCJUSBSZOVNCFSPG DPOTUSBJOUT XFOFFEUPGJOEBOBMMPDBUJPOTUSBUFHZ  GVODUJPO #&".4&"3$) / L CFBNˡ %&$0%&3*/*5 L GPS UJNFTUFQUJO/EP TDPSFT%&$0%&345&1 CFBN CFBNˡ ,#&45 TDPSFT SFUVSO CFBN<>

  5. 4UFQ(FOFSBUJOHUIFDBOEJEBUFTFU $BOEJEBUFTFU ˒ LCFTUJUFNT JF OPSNBMUPQL ˖ TJOHMFCFTUUPLFO ˠ VOGJMMFEDPOTUSBJOUT

     #FBNTJ[F L  7PDBCVMBSZTJ[F c7c  $POTUSBJOUT $  XPSET QISBTF

  6. 4UFQ"MMPDBUJOHUIFCFBN "MMPDBUFTJ[FLCFBNBDSPTT$ DPOTUSBJOUCBOLT #BOL 5IFQPSUJPOPGUIFCFBNSFTFSWFEGPSJUFNT IBWJOHNFUUIFTBNFOVNCFSPGDPOTUSBJOUT "MMPDBUJPOTUSBUFHZTFUUJOHFBDICJOTJ[FUP⌊L$⌋ "OZSFNBJOJOHTMPUTBSFBTTJHOFEUPUIFNPTUDPOTUSBJOFECBOL 

  7. &YQFSJNFOUBM4FUVQ • %BUB • 5SBJO8.5&OHMJTI(FSNBOUSBJOJOHDPSQPSB • %FW5FTUOFXTUFTU  • 1SFQSPDFTT

    • .PTFTUPLFOJ[FS • +PJOU#1&WPDBCVMBSZXJUILNFSHFPQFSBUJPO • .PEFM • MBZFS3//XJUIBUUFOUJPO • "EBNPQUJNJ[FSXJUIBCBUDITJ[FPG • &BSMZTUPQQJOH DSPTTFOUSPQZ • #BTFMJOFT • (SJE#FBN4FBSDI LC $  C   

  8. 3VOOJOHUJNF MPXFSJTCFUUFS BOE#-&6TDPSF • 5IFMJOFBSUSFOEJO$ DPOTUSBJOUT JTDMFBSGPS(#4 • 5IFDPOTUBOUUSFOEJO$ JTDMFBSGPS%#"

    • %#"JTBCPVUYTMPXFSUIBOVODPOTUSBJOFEEFDPEJOH  #-&6 #-&6 #-&6

  9. $POTUSBJOFEXPSETBSFQMBDFEJOUIFDPSSFDUQPTJUJPO 1FBSTPO`TS  

  10. $POTUSBJOFEXPSETBSFQMBDFEJOUIFDPSSFDUQPTJUJPO 

  11. .PEFMTDPSFHPFTEPXOCVU#-&6TDPSFHPFTVQ 5IJTJTBOPCTFSWBUJPOSFMBUFEUPBSFQPSU UIBU #-&6XPSTFOTJGUIFCFBNTJ[FJTNBEFUPPMBSHF  ,PFIOBOE,OPXMFT 8/.5 4JY$IBMMFOHFTGPS/FVSBM.BDIJOF5SBOTMBUJPO

  12. 'BTU-FYJDBMMZ$POTUSBJOFE%FDPEJOHXJUI%#"GPS/.5 • 5IJTTUVEZNBEFJUQPTTJCMFUPDPOUSPM/.5PVUQVU VTJOH-FYJDBMMZ$POTUSBJOFE%FDPEJOH • *O(#4 UIFDPNQVUBUJPOBMDPTUJODSFBTFEMJOFBSMZ XJUI DPOTUSBJOUT CVUJO%#"

    JUEFDSFBTFEUPDPOTUBOU PSEFS • $POTUSBJOUTBSFDPSSFDUMZQMBDFEXPBMJHONFOUJOGPSNBUJPO