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
第3回 機械学習の手法
Search
Atsushi
January 30, 2018
0
160
第3回 機械学習の手法
B3勉強会 第3回目
発表日 2018年1月30日
Atsushi
January 30, 2018
Tweet
Share
More Decks by Atsushi
See All by Atsushi
文献紹介:Automated Evaluation of Out-of-Context Errors
atsumikan
0
75
文献紹介:Correction of OCR Word Segmentation Errors in Articles from the ACL Collection through Neural Machine Translation Methods
atsumikan
0
130
文献紹介:Auxiliary Objectives for Neural Error Detection Models
atsumikan
0
66
文献紹介:Wronging a Right: Generating Better Errors to Improve Grammatical Error Detection
atsumikan
0
93
文献紹介:Low-resource OCR error detection and correction in French Clinical Texts
atsumikan
0
73
文献紹介:CMMC-BDRC Solution to the NLP-TEA-2018 Chinese Grammatical Error Diagnosis Task
atsumikan
0
92
文献紹介 : Fluency Boost Learning and Inference for Neural Grammatical Error Correction
atsumikan
0
150
文献紹介:語彙の概念化と Wikipediaを用いた英字略語の意味推定方法
atsumikan
0
120
文献紹介:The Effect of Error Rate in Artificially Generated Data for Automatic Preposition and Determiner Correction
atsumikan
0
100
Featured
See All Featured
Clear Off the Table
cherdarchuk
82
310k
[RailsConf 2023 Opening Keynote] The Magic of Rails
eileencodes
8
8.2k
Automating Front-end Workflow
addyosmani
1353
200k
Web Components: a chance to create the future
zenorocha
304
41k
Keith and Marios Guide to Fast Websites
keithpitt
407
22k
The Cult of Friendly URLs
andyhume
72
5.6k
Java REST API Framework Comparison - PWX 2021
mraible
PRO
18
6.8k
Pencils Down: Stop Designing & Start Developing
hursman
115
11k
jQuery: Nuts, Bolts and Bling
dougneiner
57
7.1k
Building Your Own Lightsaber
phodgson
97
5.6k
Infographics Made Easy
chrislema
237
17k
Six Lessons from altMBA
skipperchong
19
2.9k
Transcript
B3 130 B3
word2vec u 84)60#= 84(*% -,:; u 84$+5 (
!"/'3. u (*7 912& 84 <(*/'
word2vec !C)+ u ?5( )+$6("% u 4?5*<101. 10001.:&',8)+ u ?5*<D
@3 ;9- /+ -= -; u 7A2# ?5*<>=?5!B0
word2vec %A+- 2"/B/#//3<// #69@;2"B3<0 7$';!D)"(!) 1 C4@;V=? ,1Vid 5:
#& /( ! : (" !% , " !' , ⋯ , " !) ) .8* ! #,>+ !
word2vec ()
word2vec >>ibm query: ibm machines: 0.390068233013 digital: 0.382320284843 p&g:
0.342204213142 navigation:0.331580519676 mixte:0.323415249586 >> monday query: monday friday: 0.679558992386 late: 0.640058636665 plunge: 0.542724490166 thursday: 0.527576982975 yesterday: 0.52104818821
word2vec ]Ec u '/3!04* kb, )2HMFljsCRY '/3!04*nC O6 u
Iuyt Qb?\hH,iL:GX^@=B`o fb"-4bkb($#&)90%nC \h \hH8G;K Kp u Q&A%1&).&) 5Z>?mVSwNeHMv7 <A+/ aJg _U u [T9q rWdZ Px9j DM
word2vec - word2vec#$- u & % u &
. !/( +$' *)+$, !/( #$word2vec"
Recurrent Neural Network $+,6:K`6D EU<@ u CGN B,Y23X/ H,
L47_W2 K`F^ NN'#,1 0\9[VaS%)*,&"$- Z815;,Z81/; RNN5;T/;TPM1.(,#! J=PI >?9O= AR'#,Q]
Recurrent Neural Network RNNLM u !"(!)
u s = %& %' ⋯ %) (N! #) " ! = " %& , %' , ⋯ , %- = " %& " %' %& " %. %& %' ⋯ " %) %& %' ⋯ %)/& = "(%& ) 0 "(%1 |%& %' ⋯ %1/& ) ) 13' "(%1 |%& %' ⋯ %1/& ) ! # ! "
Recurrent Neural Network * RNNLM *! s=“+ " ”
“ ”2,%. u %./) u P(“+'" ”) u P(“+-" ”) $&(1 0 u #
Recurrent Neural Network )- ! /* %1 "+ !! $#"
" 2$ " = 1 ' ( − log- . /0 |! 2 034 D:/*'(% 0-2/4 /- ⋯ /6) $#")- ! 34( .& , !
Recurrent Neural Network #0 u 1epoch $& ,*-(/2' ,
% *s1 " 1epoch+3 u #0 .)!" 4, $&1000* .)
Recurrent Neural Network u 1epoch 208.103092604 u 2epoch 174.181462185
u 3epoch 164.917536858 u 4epoch 160.432661616 u 5epoch 157.165286105
Recurrent Neural Network u $%" Google%"RNNNeural Machine Translation
%" u u !# u ! Convolution Neural NetworkRNN
V W u word2vec u Recurrent Neural Network op u
v(2017) Chainer v2 m s n bf u 0 5 /6 gl de Nura c wiRk 9 566 2 6 6 3: 52 2 2 9: 61 62 : ) 5 6 1 6 1 71 51 6 9 u .-- ht _ .6 6 -6 2 -6 9 566 2 6 6 566 1 62 : ( 6 6 6 2 6 9