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
NLP2021 WS2 AI王 〜クイズAI日本一決定戦〜 報告スライド
Search
junya-takayama
March 19, 2021
Research
0
1k
NLP2021 WS2 AI王 〜クイズAI日本一決定戦〜 報告スライド
言語処理学会第27回年次大会ワークショップ2「AI王 〜クイズAI日本一決定戦〜」
での報告資料です
junya-takayama
March 19, 2021
Tweet
Share
More Decks by junya-takayama
See All by junya-takayama
[SNLP2021] Prefix-Tuning: Optimizing Continuous Prompts for Generation
tkym1220
0
600
Other Decks in Research
See All in Research
データxデジタルマップで拓く ミラノ発・地域共創最前線
mapconcierge4agu
0
210
EarthSynth: Generating Informative Earth Observation with Diffusion Models
satai
3
240
AIグラフィックデザインの進化:断片から統合(One Piece)へ / From Fragment to One Piece: A Survey on AI-Driven Graphic Design
shunk031
0
400
Cross-Media Information Spaces and Architectures
signer
PRO
0
240
なめらかなシステムと運用維持の終わらぬ未来 / dicomo2025_coherently_fittable_system
monochromegane
0
2.6k
Adaptive Experimental Design for Efficient Average Treatment Effect Estimation and Treatment Choice
masakat0
0
110
RHO-1: Not All Tokens Are What You Need
sansan_randd
1
170
Learning to (Learn at Test Time): RNNs with Expressive Hidden States
kurita
0
160
EcoWikiRS: Learning Ecological Representation of Satellite Images from Weak Supervision with Species Observation and Wikipedia
satai
3
110
20250502_ABEJA_論文読み会_スライド
flatton
0
200
「どう育てるか」より「どう働きたいか」〜スクラムマスターの最初の一歩〜
hirakawa51
0
840
EOGS: Gaussian Splatting for Efficient Satellite Image Photogrammetry
satai
4
480
Featured
See All Featured
Easily Structure & Communicate Ideas using Wireframe
afnizarnur
194
16k
A Tale of Four Properties
chriscoyier
160
23k
Understanding Cognitive Biases in Performance Measurement
bluesmoon
29
1.9k
How to Create Impact in a Changing Tech Landscape [PerfNow 2023]
tammyeverts
53
2.9k
Become a Pro
speakerdeck
PRO
29
5.5k
Responsive Adventures: Dirty Tricks From The Dark Corners of Front-End
smashingmag
252
21k
Optimizing for Happiness
mojombo
379
70k
Visualizing Your Data: Incorporating Mongo into Loggly Infrastructure
mongodb
48
9.7k
[RailsConf 2023 Opening Keynote] The Magic of Rails
eileencodes
30
9.6k
Cheating the UX When There Is Nothing More to Optimize - PixelPioneers
stephaniewalter
284
13k
Making Projects Easy
brettharned
117
6.4k
How To Stay Up To Date on Web Technology
chriscoyier
790
250k
Transcript
NLP2021 WS2 AIԦ ʙΫΠζAIຊҰܾఆઓʙ େൃදձ γεςϜใࠂ ͓ؾ࣋ͪղઆ 2021/03/19 େࡕେֶେֶӃใՊֶݚڀՊ ߴࢁ
൏
ࣗݾհ Ø໊લ ߴࢁ ൏ Øॴଐ େࡕେֶَ௩ݚڀࣨ % Ø5XJUUFS!ULZN Ø63-IUUQTKVOZBUBLBZBNBHJUIVCJP
ØීஈͷݚڀτϐοΫ ରγεςϜɾࣗવݴޠੜ ØࢀՃͷ͖͔͚ͬ • ࠷ۙΫΠζʹϋϚ͍ͬͯΔ͔Β • ίϯϖͱ͍͏ͷʹग़ͯΈ͔͔ͨͬͨΒ ઈࢍब׆தͰ͢ʂʂ 1
େํ • ϦʔμʔϘʔυΛҙਂ͘؍ͨ݁͠ՌͳΜ͔օͦ͏ͯͨ͠ͷͰ #&35ͱγϯϓϧͳ *3ख๏ͷΞϯαϯϒϧΛ࠾༻ ʢ·͋ײతʹදϕʔεͰdׂղ͚ͦ͏ͳײ͕͢͡Δʣ • ܭࢉࢿݯతʹ #&35ຊདྷͷઃఆతʹೖྗΛ/τʔΫϯʹ͑Δඞཁ͋Γ •
ઌ಄/τʔΫϯͱ͔Ͱͬͯɼղʹඞཁͳ͕ࣝͪΌΜͱೖΔͷ͔ʁ ʢആ༏ͷهࣄͱ͔ɼग़ԋ࡞ΘΓͱޙΖͷํʹॻ͍ͯ͋ΔΑͶʣ • ඞཁͳؚ͕ࣝ·ΕΔΑ͏ʹͪΐͬͱݡ͍ΓํΛ͍ͨ͠ 2
ઌ಄/τʔΫϯͰ͑ΒΕͳͦ͞͏ͳྫ ଉࢠʹആ༏ͷେɺ່ʹঁ༏ͷҍΛ࣋ͭɺʰϥεταϜϥΠʱ ͳͲͷөըͰ͓ͳ͡Έͷຊͷആ༏ͱ͍͑୭Ͱ͠ΐ͏ʁ ਖ਼ղهࣄɿลݠ ˠ ΫΤϦʹԠͯ͡͏·͘هࣄຊจΛཁ͍ͨ͠ʜʜ 3 ʮଉࢠʹആ༏ͷେɺ່ʹঁ༏ͷҍʯʹؔ͢Δॳग़ ʮϥεταϜϥΠʯॳग़ ઌ಄τʔΫϯʢ͍͍ͩͨʣ
ˣ·ͩ·ͩଓ͘
#&35ϕʔεछͱ *3ϕʔεछͷΞϯαϯϒϧʢॏΈ͖ͭฏۉʣ ೖྗσʔλʢڞ௨ʣ γεςϜશମ૾ 4 BERT for ཁ BERT for
લ IR (TF-IDF) *3 $IBSOHSBN จ ީิهࣄू߹ ཁث ॏ Έ ͭ ͖ ฏ ۉ ༧ଌهࣄ BSHNBY
ཁث ϞνϕʔγϣϯจதͷϑϨʔζΛଟؚ͘ΉΑ͏ʹهࣄΛཁ͍ͨ͠ ˠީิهࣄ ! ͷຊจத͔Βɼจ " தͷ୯ޠΛଟ͘ඃ෴͢ΔΑ͏ʹ จΛෳநग़͠ɼ૯୯ޠ # ҎԼͷཁจॻ
̃ ! Λ࡞͢Δ తؔɿ% = '( ∩ ' ̃ * '( ʢͨͩ͠ '( จதͷ୯ޠू߹ɼ' ̃ * ཁจॻதͷ୯ޠू߹ʣ % ྼϞδϡϥੑΛ࣋ͭͨΊɼ্࣮ % ͕࠷େ͖͘ͳΔจΛஞ࣍తʹ ̃ ! ʹՃ͍͑ͯ͘ΞϓϩʔνΛͱΔʢᩦཉ๏ʣ 5
ཁثͷग़ྗྫ จ ଉࢠʹആ༏ͷେɺ່ʹঁ༏ͷҍΛ࣋ͭɺʰϥεταϜϥΠʱͳͲ ͷөըͰ͓ͳ͡Έͷຊͷആ༏ͱ͍͑୭Ͱ͠ΐ͏ʁ ਖ਼ղهࣄʢลݠʣݪจลݠʢΘͨͳ͚Μɺ݄ʣɺຊͷആ༏ɻຊ໊ಉ͡ɻ৽ ׁݝڕপ܊ਆଜʢݱɿڕপࢢʣग़ɻԋܶूஂԁΛܦ͔ͯΒέΠμογϡॴଐɻੈք֤ࠃʹ͓͍ͯөըΛத ৺ʹςϨϏυϥϚɺɺςϨϏίϚʔγϟϧͱ෯͘׆༂͍ͯ͠ΔຊΛද͢Δആ༏ͷҰਓɻDNɺମॏ LHɻͷล྄ҰըՈͱͯ͠׆ಈ͍ͯ͠Δɻ৽ׁݝڕপ܊ਆଜʹͯڞʹڭࢣΛ͍ͯͨ྆͠ͷݩʹੜ·ΕΔɻ ྆ͷసۈͰ༮গظΛೖଜɺकଜʢͱʹڕপࢢʣɺߴాࢢʢ্ӽࢢʣͰա͢͝ɻʜʜʢதུʣʜʜҰ༂શࠃతͳ ਓؾΛ֫ಘɺελʔμϜʹͷ্͕͠Δɻ·ͨɺͦͷࠒ͔ΒՎखͱͯ͠ࠒ·Ͱ׆ಈ͍ͯͨ͠ɻҎ߱ɺɾςϨ
ϏυϥϚͳͲͰ࣍ʑͱେΛԋ͡ɺલ్༸ʑʹݟ͑ͨɺөըॳओԋͱͳΔͣͰ͋ͬͨʰఱͱʢ୯ޠʣ ਖ਼ղهࣄʢลݠʣཁ ลݠʢΘͨͳ͚Μɺ݄ʣɺຊͷആ༏ɻຊ໊ಉ͡ɻ৽ ׁݝڕপ܊ਆଜʢݱɿڕপࢢʣग़ɻຊࠃ֎өըॳग़ԋͱͳͬͨΞϝϦΧөըʰϥεταϜϥΠʱ ʢެ։ʣͰɺลಉͷୈճΞΧσϛʔॿԋஉ༏ͳΒͼʹୈճΰʔϧσϯάϩʔϒॿԋஉ༏ɺ ୈճαλʔϯॿԋஉ༏ʹϊϛωʔτ͞ΕΔߴ͍ධՁΛಘΔɻ·ͨɺөըެ։ͱಉ࣌͡ظʹൃදͨࣗ͠Βͷஶॻ ʰ୭ 8)0".* ʱͰɺ͔ͭͯന݂පͷ࣏ྍதසൟʹड͚ͨ༌݂ʢओʹ݂খ൘༌݂ʣ͕ݪҼͰ$ܕ؊ԌΠϧεʹײછ͠ɺ ʰ໌ͷهԱʱͷࡱӨͦͷ࣏ྍͷ෭࡞༻ʹ·͞Εͳ͕Βߦ͍ͯͨ͜͠ͱΛࠂനɻ࣌Λಉͯ͘͡͠ςϨϏ౦ژͷαε ϖϯευϥϚͷڞԋΛػʹΓ߹ͬͨঁ༏ͷೆՌาͱຊ֨తʹަࡍΛ։࢝͠ɺಉ݄ʹ࠶ࠗɻͳ͓ɺଉࢠͷେ ͱؒతʹͰ͋Δ͕ڞԋྺ͋Δ͕ɺ່ͷҍͱऀۀҎ֎Ͱڞԋͨ͜͠ͱͳ͍ɻ 6
#&35ϕʔεྨث ͋Δબࢶ͕ਖ਼ղ͔Ͳ͏͔ఆ͢Δࡍʹଞͷબࢶߟྀ͍ͤͨ͞ ˠ #&35 4FMG"UUFOUJPO-BZFSͷ֊ܕΞʔΩςΫνϟΛ࠾༻ 7 ࠷ऴ <$-4> ࠷ऴ .BY1PPMJOH
<$-4>จ<4&1>هࣄ<4&1> <$-4>จ<4&1>هࣄ<4&1> ʜ ʜ BERT BERT BERT Self Attention Layer Softmax Linear Linear Linear
*3Ϟσϧ <5'*%'ϕʔε> • จͷ 5'*%'ϕΫτϧͱީิهࣄͷ 5'*%'ϕΫτϧͷ DPTྨࣅ͕ߴ͍هࣄΛਖ਼ղީิͱ͢ΔγϯϓϧͳϞσϧ • ͨͩ͠ *%'ʢίʔύεશମͰͳ͘ʣ֤͝ͱʹ
ͦͷͷީิهࣄશମʢ݅ʣ͔Βܭࢉ <ཧ༝>ީิهࣄू߹ͦͦʢ8JLJQFEJB7FDతʹʣྨࣅ͓ͯ͠Γɼ ίʔύεશମ͔Βܭࢉͨ͠ *%'Λ༻͍Δͱ 5'*%'ϕΫτϧ͕௵Εͦ͏ ʢͳؾ͕͢Δʂʂʣʢະݕূʣ <$IBSBDUFS/HSBNϕʔε> • จͷ /HSBNू߹ͱީิهࣄͷ /HSBNू߹ͷ 4JNQTPO • ୯ޠΑΓจࣈ /HSBNͷํ͕ "DDVSBDZ͕͘Β͍ߴ͔ͬͨ 8
ͦͷଞࡉʑͱͨ͠ʢCVUΫϦςΟΧϧͳʣલॲཧ • <*3 $IBS >ίʔύεதͷස্Ґޠͷ͏ͪʮετοϓϫʔυͳʔʯͱ ࢥͬͨͷΛετοϓϫʔυϦετʹՃɽείΞܭࢉ࣌ʹআ֎ • <*3 ྆ํ >ΤϯςΟςΟ໊͕จதʹؚ·Ε͍ͯͨΒਖ਼ղީิ͔Βআ
ʢʮIPHF GVHB ͱ͋ͱԿͰ͠ΐ͏ʁʯͰ IPHF GVHB ͕બΕ͕ͪ ͩͬͨͨΊʣ • <ཁث>ετοϓϫʔυతؔ ! ͷܭࢉ࣌ʹߟྀ͠ͳ͍ • <ཁث>ɻͰจׂ͢Δ͕ɼ͗͢Δ߹ʢʣ૭෯Λ ୯ޠ ͱͯ͠ɼ૭Λٖࣅతͳจͱ͢Δ • <ཁث>લจ࠷ॳ͔ΒཁจʹՃ 9
%FWͰͷ࣮ݧ݁Ռ <ओཁͳ࣮ݧઃఆ> • #&35Ϟσϧͷ࠷େτʔΫϯɿʢϞσϧڞ௨ʣ • #&35ࣄલֶशࡁΈϞσϧɿcl-tohoku/bert-base-japanese-whole-word-masking • *3 $IBSBDUFS/HSBN ͷ
A/A • ܇࿅σʔλɿ5SBJOͷΈɽΞϯαϯϒϧͷॏΈ %FWͰௐ <࣮ݧ݁Ռ> 10 Ϟσϧ "DDVSBDZ<> %FW %FW *3 5'*%' 64.72 61.79 *3 $IBSCJHSBN 72.66 69.71 *3 $IBSUSJHSBN 74.77 73.82 #&35 લτʔΫϯ 84.62 83.55 #&35 ཁ 88.94 89.67 Ξϯαϯϒϧ 92.05 91.14
ϦʔμʔϘʔυ "DDॱҐҐλΠʢ࣌ʣ Ґ ʢ࣌ʣ ʢʮ·͋ҐҎʹΔΖʯͱ͔ࢥͬͯͨͷʹʜʜʣ 11
ॴײ <ল> • ʮͰ͕͢ʯͷʮͰ͕͢ʯલ෦ͱ͔ฒྻͷྻڍ෦ͱ͔ɼ هࣄݕࡧʹ͍Βͳͦ͏ͳ෦Λؤுͬͯআڈͯ͠ΈΔ͖͔ͩͬͨ • จͱީิهࣄͷؒʹ͏ ϗοϓ͘Β͍ඞཁͦ͏ͳ͕݁ߏ͕͋ͬͨɼ ݟͯݟ͵ৼΓΛͯ͠͠·ͬͨ <ײ>
• ࠷ۙ #&35 #"35ʹͱʹ͔͘ͳΜͰಥͬࠐΉ͜ͱ͕ଟ͔ͬͨͷͰɼ ٱʑʹࣗવݴޠॲཧಓͰటष͍લॲཧΛΕָ͔ͯͬͨ͠Ͱ͢ • ίϯϖָ͍͠Ͱ͢Ͷɽ,BHHMFͱ͔ͬͯΈΑ͏ͱࢥ͍·ͨ͠ ओ࠵ऀͷօ༷ɼָ͍͠ίϯϖΛاըͯͩ͘͠͞Γ͋Γ͕ͱ͏͍͟͝·ͨ͠ʂʂʂ ઈࢍब׆தͰ͢ʂʂ 12