Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Speaker Deck
PRO
Sign in
Sign up for free
(Reading )Does BERT Know that the IS-A Relation Is Transitive?
ando
January 23, 2023
Science
0
5
(Reading )Does BERT Know that the IS-A Relation Is Transitive?
Komachi lab
ando
January 23, 2023
Tweet
Share
More Decks by ando
See All by ando
博士論文公聴会資料
ando55
0
12
Is In-hospital Meta-information Useful for Abstractive Discharge Summary Generation?
ando55
0
9
(読み会)Evaluating Factuality in Text Simplification
ando55
0
12
(Reading) Relational Multi-Task Learning Modeling Relations between Data and Tasks
ando55
0
48
(Reading )Agreeing to Disagree: Annotating Offensive Language Datasets with Annotators’ Disagreement
ando55
0
35
(Reading) Preregistering NLP research
ando55
0
38
(Reading) Predictive Adversarial Learning from Positive and Unlabeled Data
ando55
0
79
Argument Invention from First Principles
ando55
2
170
Other Decks in Science
See All in Science
Causal Impact -paper summary-
hoxomaxwell
1
120
20221005_AI勉強会
ecoopnet
0
240
SpaceXから学ぶ人生観 ~元JWが自分の人生を取り戻すヒント~
caesar2022
0
180
標本調査法の基礎と実践
arumakan
0
310
Tokyo.R RStudioでグラフをちょっときれいに出力する - CairoとAGG -
bob3bob3
0
870
論文紹介: Transformer Memory as a Differentiable Search Index (NeurIPS 2022)
ynakano
3
390
古典的な時系列解析フレームワークの理論とその実装 / theory-and-implementation-of-classical-time-series-analysis-frameworks
saburoku
0
110
iPhoneと一緒なら 一人でも頑張れる 野外調査の3次元記録
ishiijunpei
0
370
調整変数の選び方
koro485
2
3.5k
ESG評価に対する自然言語処理の活用Workshop
icoxfog417
0
140
SHINOMIYA Nariyoshi
genomethica
0
330
Successful social norms with ternary reputation for indirect reciprocity
yohm
0
130
Featured
See All Featured
Large-scale JavaScript Application Architecture
addyosmani
499
110k
JazzCon 2018 Closing Keynote - Leadership for the Reluctant Leader
reverentgeek
175
9.1k
How GitHub (no longer) Works
holman
298
140k
"I'm Feeling Lucky" - Building Great Search Experiences for Today's Users (#IAC19)
danielanewman
217
21k
Building a Modern Day E-commerce SEO Strategy
aleyda
6
4.5k
GraphQLの誤解/rethinking-graphql
sonatard
39
7.8k
Creating an realtime collaboration tool: Agile Flush - .NET Oxford
marcduiker
6
840
Understanding Cognitive Biases in Performance Measurement
bluesmoon
2
390
The Art of Programming - Codeland 2020
erikaheidi
35
11k
Visualizing Your Data: Incorporating Mongo into Loggly Infrastructure
mongodb
32
6.7k
Side Projects
sachag
451
37k
Embracing the Ebb and Flow
colly
75
3.6k
Transcript
ACL2022 Short paper
- #&35͕*4"ؔͷ5SBOTJUJWFੑ࣭ͱͲͷఔҰக͢Δ͔ΛఆྔԽɻ - #&35ͷ༧ଌ*4"ؔͷਪҠੑʹશʹैΘͳ͍͜ͱ͕໌Β͔ʹͳͬͨ 5SBOTJUJWF A is-a B & →
A is-a C︖ B is-a C - #&35*4"ؔΛͷ"DDVSBDZͰ༧ଌͰ͖Δ - #&35͕"JTB#ͱ#JTB$Λਖ਼͘͠༧ଌͨ͠߹ɺ"JTB$ ˠͦͷ··ࣝϕʔεͱͯ͠ѻ͑ͳ͍ 2 概要
- ໊ࢺΛ༻͍Δ 3 Task definition 語義のペア Is-aであれば1, でなければ0
-\TcT㱨4 BOEIZQP T 㱵 ^ 8PSE/FUͷશͯͷϦʔϑηϯεɻ 48PSE/FUͷશͯͷޠٛͷू߹ɺIZQP T ҙຯTͷԼҐޠٛͷू߹Λද͢ɻ -
*4"ΛऔΓग़͢ • -͔ΒMFBGͷޠٛ/ΛϥϯμϜʹநग़͠ɺ/ΛSPPU3ʹଓͯ͠ɺύεQ /ˠ3 Λಘ Δ Q͔Β"ɺ#ɺ$ͷͭͷޠٛΛϥϯμϜʹநग़͠ɺ"ɺ#ɺ$ͷޠٛͷྫจΛಘΔ ͜ͷ݁Ռɺλϓϧʢ"ɺ#ɺ$ʣͱͭͷਖ਼ྫʢ"ɺ#ʣɺʢ#ɺ$ʣɺʢ"ɺ$ʣ͕ಘ ΒΕΔɻ - /PU*4"ΛऔΓग़͢ • -͔Βೋͭͷ/ /ΛϥϯμϜʹநग़͠ɺೋͭͷύεQ /ˠ3 ͱQ /ˠ3 Λ ಘΔQɺQ͔ΒͦΕͧΕ"⒌ɺ#⒌ΛϥϯμϜʹαϯϓϦϯά͠ɺޠٛ"⒌ɺ#⒌ͷྫจΛಘ Δ "⒌͕#⒌ˠ3ͷύε্ʹͳ͍߹ɺ·ͨͦͷٯͷ߹ɺෛྫʢ"⒌ #⒌ʣΛಘΔ - σʔληοτͰɺݸͷλϓϧ͕͋Γɺݸͷਖ਼ྫͱݸͷෛྫ͕͋Δɻ 4 Positive, Negative example抽出
- #&35ͷ࠷ऴͷϕΫτϧΛ༻ʢTVCXPSEBWFSBHFʣ - #&35ϑΝΠϯνϡʔχϯά͠ͳ͍ɺ࠷ۙʢOOʣྨثΛ༻ ˠγϯϓϧͳͷΛ͍͍ͨ - ϢʔΫϦουڑΛڑࢦඪͱͯ͠༻͍Δɻ 5 分類 語義2
語義1
- ςετηοτͰɺ " # ͱ # $ ͷฏۉϗοϓͦΕͧΕ ͱ Ͱ͋ͬͨɻ
- ͜ͷࠩɺ8PSE/FUͷྫจΛ࣋ͭޠ͕ٛۉʹ͍ͯ͠ͳ͍͜ͱʹىҼ͍ͯ͠Δɻ ʢ্ҐޠྫจΛ͍͍࣋ͬͯ͢ʣ 6 Hop数
- ϗοϓ͕গͳ͍߹ʢʣɺϗοϓ͕ଟ͍߹ʢʣΑΓߴ͍ਫ਼Ͱ*4"ؔ Λ༧ଌͰ͖Δ - ϗοϓ͕૿Ճͯ͠༧ଌਫ਼େ͖͘ԼͤͣɺϗοϓͰਫ਼ʹେ͖ͳΒ͖ͭ ͳ͍ɻ 7 実験結果 ?
- #&35͕ʢ"ɺ#ʣɺʢ#ɺ$ʣΛਖ਼͘͠༧ଌͨ͠߹ʹɺ*4"ؔʢ"ɺ$ʣΛਖ਼͘͠༧ଌ ͢ΔසΛଌఆ - Q "# ͷ " # ͷ͏ͪɺਖ਼͘͠༧ଌ͞Εͨ
" # ͷׂ߹ - Q "$c"# #$ ɺ " # ͱ # $ ͕ਖ਼͘͠༧ଌ͞ΕΔ߹ɺ " $ ͕ਖ਼͘͠༧ଌ͞ΕΔׂ߹ - "JTB$͕ৗʹਖ਼͘͠༧ଌ͞Εͳ͍͜ͱɺ#&35͕ཧతʹҰ؏ͨ͠༧ଌΛ͢ΔೳྗΛ ͍͍ܽͯΔ͜ͱΛ͍ࣔࠦͯ͠Δɻ 8 Prediction Ability for Transitivity