Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Sign up for free
Menu
Search
Features
All features
Private URLs
Password Protection
Custom URLS
Scheduled publishing
Remove Branding
Restrict embedding
Deck Collections
Notes
Features
All features
Private URLs
Password Protection
Custom URLS
Scheduled publishing
Remove Branding
Restrict embedding
Deck Collections
Notes
Explore
Featured decks
Featured speakers
Programming
Technology
Storyboards
Explore
Featured decks
Featured speakers
Programming
Technology
Storyboards
Pricing
Search
Sign in
Sign up for free
【輪講資料】Zero-shot Cross-lingual Semantic Parsing
Search
Yano
May 03, 2023
200
0
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
【輪講資料】Zero-shot Cross-lingual Semantic Parsing
研究室内の輪講で使った資料です。
Yano
May 03, 2023
More Decks by Yano
See All by Yano
【輪講資料】Lost in Execution: On the Multilingual Robustness of Tool Calling in Large Language Models
yano0
0
42
【輪講資料】How much can language models memorize?
yano0
0
250
【輪講資料】Decoding Dense Embeddings: Sparse Autoencoders for Interpreting and Discretizing Dense Retrieval
yano0
0
170
【輪講資料】Length-Induced Embedding Collapse in PLM-based Models
yano0
0
190
【輪講資料】How Do Large Language Models Acquire Factual Knowledge During Pretraining?
yano0
0
390
NLP2025参加報告
yano0
0
770
【輪講資料】ReAct: Synergizing Reasoning and Acting in Language Models / Tree of Thoughts: Deliberate Problem Solving with Large Language Models
yano0
0
300
【輪講資料】SimLM: Pre-training with Representation Bottleneck for Dense Passage Retrieval
yano0
2
420
【輪講資料】From Zero to Hero: On the Limitations of Zero-Shot Language Transfer with Multilingual Transformers
yano0
0
170
Featured
See All Featured
SERP Conf. Vienna - Web Accessibility: Optimizing for Inclusivity and SEO
sarafernandez
2
1.6k
Prompt Engineering for Job Search
mfonobong
0
460
B2B Lead Gen: Tactics, Traps & Triumph
marketingsoph
0
250
Lightning Talk: Beautiful Slides for Beginners
inesmontani
PRO
2
700
The untapped power of vector embeddings
frankvandijk
2
1.9k
Efficient Content Optimization with Google Search Console & Apps Script
katarinadahlin
PRO
1
880
We Are The Robots
honzajavorek
0
380
Discover your Explorer Soul
emna__ayadi
2
1.3k
実際に使うSQLの書き方 徹底解説 / pgcon21j-tutorial
soudai
PRO
203
76k
Cheating the UX When There Is Nothing More to Optimize - PixelPioneers
stephaniewalter
287
14k
A Modern Web Designer's Workflow
chriscoyier
699
190k
The #1 spot is gone: here's how to win anyway
tamaranovitovic
4
1.2k
Transcript
Zero-shot Cross-lingual Semantic Parsing Tom Sherborne, Mirella Lapata ACL 2022
֓ཁ 2 ✓ θϩγϣοτଟݴޠҙຯղੳϞσϧɺZX-parseͷఏҊ • ݴޠؒજࡏදݱͷΞϥΠϝϯτʹண͠సҠֶशͷޡࠩΛ࠷খԽ • ෳͷଛࣦؔΛಋೖ͢ΔϚϧνλεΫֶशʹΑΓҟͳΔݴޠͷ જࡏදݱ͕ྨࣅ͢ΔΑ͏ʹ •
ରݴޠͷϖΞσʔλΛඞཁͱͤͣɺӳޠϖΞͱରݴޠͷࣗવ จͷΈར༻ ✓ θϩγϣοτҙຯղੳλεΫʹ͓͍ͯෳͷݴޠͰߴ͍ੑೳ
બΜͩཧ༝ • ݴޠԣஅతͳݚڀʹڵຯ͕͋ΔͨΊ • ಛʹݴޠԣஅతͳજࡏۭؒΛ࡞ΔͰ໘നͦ͏ • ࣗͷݚڀΛؚΉ͞·͟·ͳసҠֶशλεΫͰར༻Ͱ͖ͦ͏ 3
ҙຯղੳ (Semantic Parsing) • ࣗવݴޠͷൃΛཧܗࣜ(Logical Form)ʹม • ͞·͟·ͳλεΫͰॏཁͳΠϯϑϥ • ࣭ԠɺରγεςϜɺػցͷࢦࣔ…
4 -JTU fl JHIUTGSPN4BO'SBODJTDPUP1JUUTCVSHI 4&-&$5%*45*/$5 fl JHIU@ fl JHIU@JE'30.ʜ
• ෳͷݴޠ͔ΒͳΔࣗવจΛಉ͡ཧܗࣜʹม[1] ଟݴޠҙຯղੳ (Cross-lingual Semantic Parsing) 5 <>.VMUJMJOHVBM4FNBOUJD1BSTJOH1BSTJOH.VMUJQMF-BOHVBHFTJOUP4FNBOUJD3FQSFTFOUBUJPOT 4&-&$5%*45*/$5 fl
JHIU@ fl JHIU@JE'30.ʜ -JTU fl JHIUTGSPN4BO 'SBODJTDPUP1JUUTCVSHI αϯϑϥϯγείൃϐοπ όʔάߦ͖ͷϑϥΠτΛ ϦετΞοϓ͍ͯͩ͘͠͞ɻ ʹ
ઌߦݚڀ • ӳޠͷࣗવจ-ཧܗࣜͷฒྻσʔλΛ༁͠ɺ֤ରݴޠͷࣗવ จ-ཧܗࣜϖΞσʔλΛར༻ • ػց༁Λ༻͍Δ߹[2] • ػց༁ᘳͰͳ͍ • ಛʹϦιʔεݴޠʹ͓͍ͯ
ߴ࣭ͳػց༁͍͠ • ਓख༁Λ༻͍Δ߹[3] • ߴίετ 6 <>#PPUTUSBQQJOHB$SPTTMJOHVBM4FNBOUJD1BSTFS <>/FVSBM"SDIJUFDUVSFTGPS.VMUJMJOHVBM4FNBOUJD1BSTJOH ཧܗࣜ ӳޠ ରݴޠ ༁
• ରݴޠͷฒྻσʔλΛΘͳ͍ θϩγϣοτͱ͢Δ ➡ ӳޠͷࣗવจ-ཧܗࣜϖΞσʔλͱ ରݴޠͷࣗવจͷΈར༻ ✓ Ϟσϧߏ • ӳޠ͔Βಘͨજࡏදݱ͔Β
ཧܗࣜΛੜ͢Δσίʔμ • ݴޠԣஅతͳʢӳޠͱྨࣅͨ͠ʣ જࡏදݱΛ֫ಘ͢ΔΤϯίʔμ ఏҊख๏ 7 αϯϑϥϯγείൃϐοπ όʔάߦ͖ͷϑϥΠτΛ ϦετΞοϓ͍ͯͩ͘͠͞ɻ 4&-&$54&-&$54&-&$5 8)&3& '30.4&-&$5 Τϯίʔμ σίʔμ ❌ 4&-&$5%*45*/$5 fl JHIU@ fl JHIU@JE'30.ʜ -JTU fl JHIUTGSPN4BO 'SBODJTDPUP1JUUTCVSHI ⭕ Τϯίʔμ σίʔμ ˛ΞϥΠϝϯτΛߦΘͳ͍߹ɺݴޠ͝ͱʹ ҟͳΔજࡏදݱͱͳΓग़ྗ͕ҟͳΔ જࡏදݱ
ϚϧνλεΫֶश • ̍ͭͷΤϯίʔμʹର͠ෳͷతؔΛ༻͍ɺಉ࣌ʹ࠷దԽ[4] • ଟݴޠλεΫʹ͓͍ͯɺʮιʔεݴޠ (ӳޠͳͲ) ͰͷඪλεΫʴ ιʔεݴޠͱରݴޠͷΞϥΠϝϯτʯΛಉ࣌ʹ࠷దԽ͢Δݚڀ ͕ଘࡏ •
ԻݴޠཧղɺςΩετ؆ུԽɺґଘੑߏจղੳɺػց༁ 8 <>.VMUJUBTL4FRVFODFUP4FRVFODF-FBSOJOH
ఏҊϞσϧɿZX-Parse • ϚϧνλεΫSeq2seqϞσϧ ✓ తͷλεΫʹՃ͑ิॿతͳλεΫΛಋೖ • DLF ɿཧܗࣜͷੜ • DNL
ɿࣗવݴޠͷੜ • LPɿݴޠ༧ଌ 9 Τϯίʔμ z DLF LP ࣗવจ DNL ֶश(EN) ਪ જࡏදݱ Transformer x 6 (mBARTͷΤϯίʔμ) Transformer x 6 ֶश(ଞݴޠ) ཧܗࣜͷੜ ӳޠͰͷΈֶश
DLF ɿ Generating Logical Forms 10 Τϯίʔμ z DLF LP
List fl ights from San Francisco to Pittsburgh? DNL SELECT DISTINCT fl ight_1. fl ight_id FROM … • ӳޠͷࣗવจ͔Βಉ͡ҙຯͷཧܗࣜΛੜ͢Δ ➡ཧܗࣜੜೳྗΛʹ͚ͭΔ ֶश࣌ ਪ࣌ જࡏදݱ
• ରݴޠͷࣗવจʹϊΠζΛՃ͑ɺ࠶ߏங͘͠ӳޠ༁ ➡֤ݴޠͷࣗવจʹΤϯίʔμʔΛదԠͤ͞Δ • ݴޠݻ༗ͷಛੑΛʹ͚Δ DNL ɿ Generating Natural Language
11 Τϯίʔμ z DLF LP αϯϑϥϯγείൃϐοπ όʔάߦ͖ͷϑϥΠτΛ ϦετΞοϓ͍ͯͩ͘͠͞ɻ DNL Н ϊΠζ Н αϯϑϥϯγείൃϐοπ όʔάߦ͖ͷϑϥΠτΛ ϦετΞοϓ͍ͯͩ͘͠͞ɻ List fl ights from San Francisco to Pittsburgh? ֶश࣌ ਪ࣌ જࡏදݱ
LPɿ Language Prediction 12 Τϯίʔμ z DLF LP List fl
ights from San Francisco to Pittsburgh? DNL English ֶश࣌ ਪ࣌ • ೖྗจͷݴޠΛྨثͰ༧ଌ͢Δ ➡ΤϯίʔμʹݴޠΛ۠ผͤ͞Δ જࡏදݱ
LPɿ Language Prediction 13 Τϯίʔμ z DLF LP List fl
ights from San Francisco to Pittsburgh? DNL English ֶश࣌ ਪ࣌ • ೖྗจͷݴޠΛྨثͰ༧ଌ͢Δ ➡ΤϯίʔμʹݴޠΛ۠ผͤ͞Δ ✦ ٯ࣌ʹޯΛసͤ͞Δ ➡ ݴޠΛ۠ผͤ͞ͳ͍ • ݴޠʹͱΒΘΕͳ͍දݱ જࡏදݱ
• ಉ࣌ʹ̏ͭͷతؔΛ࠷దԽ • ཧܗࣜͷੜʴݴޠ͝ͱͷಛΛֶशʴݴޠΛ۠ผ͠ͳ͍ • ରݴޠͷϖΞίʔύεΛඞཁͱ͠ͳ͍θϩγϣοτҙຯղੳ ZX-Parse 14 Τϯίʔμ z
DLF LP DNL ٯ࣌ ཧܗࣜͷੜ ࣗવݴޠͷੜ ʢ࠶ߏ/༁ʣ ݴޠ༧ଌ − ∂LLP ∂θ ∂LNL ∂θ ∂LLF ∂θ જࡏදݱ
σʔληοτ ✦ ҙຯղੳσʔληοτ • ATIS • ཱྀߦυϝΠϯͷଟݴޠࣗવจͱཧܗࣜͷϖΞίʔύε • ӳޠ, ϑϥϯεޠ,
ϙϧτΨϧޠ, εϖΠϯޠ, υΠπޠ, தࠃޠ, ώϯσΟʔޠ, τϧίޠ • Overnight • ̔υϝΠϯͷӳޠࣗવจͱཧܗࣜͷϖΞίʔύε • ਪ࣌ͷΈ༁Ͱ࡞͞ΕͨதࠃޠͱυΠπޠσʔλΛར༻ ✦ ࣗવݴޠσʔληοτ • MKQA • ࣭จͷର༁ίʔύε • ӳޠ, ϑϥϯεޠ, ϙϧτΨϧޠ, εϖΠϯޠ, υΠπޠ, தࠃޠ • ParaCrawl • Webର༁ίʔύε 15
σʔληοτ ✦ ҙຯղੳσʔληοτ • ATIS • ཱྀߦυϝΠϯͷଟݴޠࣗવจͱཧܗࣜͷϖΞίʔύε • ӳޠ, ϑϥϯεޠ,
ϙϧτΨϧޠ, εϖΠϯޠ, υΠπޠ, தࠃޠ, ώϯσΟʔޠ, τϧίޠ • Overnight • ̔υϝΠϯͷӳޠࣗવจͱཧܗࣜͷϖΞίʔύε • ਪ࣌ͷΈ༁Ͱ࡞͞ΕͨதࠃޠͱυΠπޠσʔλΛར༻ ✦ ࣗવݴޠσʔληοτ • MKQA • ࣭จͷର༁ίʔύε • ӳޠ, ϑϥϯεޠ, ϙϧτΨϧޠ, εϖΠϯޠ, υΠπޠ, தࠃޠ • ParaCrawl • Webର༁ίʔύε 16
༁ϕʔεϥΠϯ ✦ “࠷খݶͷྗ”ϕʔεϥΠϯ • ػց༁Λ༻͍ͯ࡞ͨ͠ϕʔεϥΠϯ • Translate-Test • ςετηοτΛӳޠʹ༁ •
Translate-Train • ֶशηοτΛରݴޠʹ༁ ✦ “࠷େݶͷྗ”ϕʔεϥΠϯ • Monolingual Training • ֶशηοτΛਓखͰରݴޠʹ༁ 17
࣮ݧ݁Ռɿ༁ϕʔεϥΠϯͱͷൺֱ • ӳޠҎ֎ͷݴޠͷθϩγϣοτҙຯղੳλεΫͰSOTA 18 • Monolingual Training: ֶशηοτΛਓखͰରݴޠʹ༁ • Translate-Train:
ֶशηοτΛରݴޠʹ༁ • Translate-Test: ςετηοτΛӳޠʹ༁ • ATIS: ࣭จͱཧܗࣜͷϖΞσʔλ • Overnight: ෳυϝΠϯͷจͱཧܗࣜͷϖΞσʔλ ※ώϯσΟʔޠ(HI)ɺτϧίޠ(TR)ର༁ίʔύε͕ ଘࡏ͠ͳ͍ͨΊิॿతͰͷֶशʹؚ·Ε͍ͯͳ͍
࣮ݧ݁Ռɿ࠷খݶͷྗϕʔεϥΠϯͱͷൺֱ • ิॿతͰͷֶशΛߦͳͬͨશͯͷݴޠ(HI,TRҎ֎)ʹ͓͍ͯੑೳ্ • OvernightATIS΄Ͳੑೳ্͍ͯ͠ͳ͍ • υϝΠϯ͕ଟ༷ͳͨΊɺ՝͕ෳ߹తʹͳ͍ͬͯΔ • ӳޠʹ͍ۙݴޠͰTranslate-TrainTranslate-TestΛԼճΔ͕ɺԕ͍ݴޠ Ͱੑೳ͕Լ͢Δ
• ػց༁ͷํͷӨڹʁ 19
࣮ݧ݁ՌɿαϒλεΫͷӨڹ 20 • LFͷΈͷ߹ɺTranslate-TestΑΓੑೳ͕͍ • NLͱLPͲͪΒ͔Ճ͢Δ࣌ɺNLΛՃͨ͠ํ͕ੑೳ͕ߴ͍ • ֤ݴޠͷࣗવݴޠจʹదԠ͢Δํ͕େࣄ
࣮ݧ݁ՌɿαϒλεΫͷӨڹ 21 • LFͷΈͷ߹ɺTranslate-TestΑΓੑೳ͕͍ • NLͱLPͲͪΒ͔Ճ͢Δ࣌ɺNLΛՃͨ͠ํ͕ੑೳ͕ߴ͍ • ֤ݴޠͷࣗવจʹదԠ͢Δ͜ͱ͕େࣄ
࣮ݧ݁ՌɿαϒλεΫͷӨڹ • αϒλεΫͰֶशʹؚ·Εͳ͍ݴޠ(HI,TR)Ͱੑೳ্ • ֶश͍ͯ͠ͳͯ͘ݴޠؒͷજࡏදݱ͕վળ͞Ε͍ͯΔ 22
࣮ݧ݁ՌɿαϒλεΫͷӨڹ • αϒλεΫͰֶशʹؚ·Εͳ͍ݴޠ(HI,TR)Ͱੑೳ্ • ֶश͍ͯ͠ͳͯ͘ݴޠؒͷજࡏදݱ͕վળ͞Ε͍ͯΔ 23 ˝ຒΊࠐΈͷՄࢹԽ
࣮ݧ݁ՌɿαϒλεΫͷӨڹ • ୯ݴޠͷΈར༻͢Δ߹(τ = 0.0)ΑΓ༁ߦ͏߹(τ = 0.5) ͷੑೳ͕ߴ͍ • దͳจΛར༻͢Δ߹(ParaCrawl)ΑΓਪ࣌ͷೖྗͱಉ࣭͡
จΛֶशʹར༻͢Δ߹(MKQA)ͷੑೳ͕ߴ͍ 24
·ͱΊ ✓ θϩγϣοτଟݴޠҙຯղੳϞσϧɺZX-parseͷఏҊ • જࡏදݱͷΞϥΠϝϯτʹண͠సҠֶशͷޡࠩΛ࠷খԽ • ෳͷଛࣦؔΛಋೖ͢ΔϚϧνλεΫֶशʹΑΓҟͳΔݴޠͷ જࡏදݱ͕ྨࣅ͢ΔΑ͏ʹ • ରݴޠͷϖΞσʔλΛඞཁͱͤͣɺӳޠϖΞͱରݴޠͷࣗવ
จͷΈར༻ ✓ θϩγϣοτҙຯղੳλεΫʹ͓͍ͯෳͷݴޠͰߴ͍ੑೳ 25