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
Kaggle Google Quest Q&A Labeling - 23th place s...
Search
Shuhei Goda
February 28, 2020
Technology
4
4.1k
Kaggle Google Quest Q&A Labeling - 23th place solution
Shuhei Goda
February 28, 2020
Tweet
Share
More Decks by Shuhei Goda
See All by Shuhei Goda
Turing × atmaCup #18 - 1st Place Solution
hakubishin3
0
970
ジョブマッチングサービスにおける相互推薦システムの応用事例と課題
hakubishin3
3
1k
とある事業会社にとっての Kaggler の魅力
hakubishin3
8
2.8k
課題の解像度が荒かったことで意図した改善ができなかった話
hakubishin3
3
1k
Wantedly におけるマッチング体験を最大化させるための推薦システム
hakubishin3
4
1.3k
Recommendation Industry Talks #1 Opening
hakubishin3
1
420
会社訪問アプリ「Wantedly Visit」での シゴトに関する興味選択機能と推薦改善
hakubishin3
0
680
論文紹介: Improving Implicit Feedback-Based Recommendation through Multi-Behavior Alignment(Xin Xin et al., 2023)
hakubishin3
0
670
Feedback Prize - English Language Learning における擬似ラベルの品質向上の取り組み
hakubishin3
0
1.1k
Other Decks in Technology
See All in Technology
フルカイテン株式会社 エンジニア向け採用資料
fullkaiten
0
9.1k
Shirankedo NOCで見えてきたeduroam/OpenRoaming運用ノウハウと課題 - BAKUCHIKU BANBAN #2
marokiki
0
180
Simplifying Cloud Native app testing across environments with Dapr and Microcks
salaboy
0
130
ACA でMAGI システムを社内で展開しようとした話
mappie_kochi
1
310
OCI Network Firewall 概要
oracle4engineer
PRO
2
7.8k
小学4年生夏休みの自由研究「ぼくと Copilot エージェント」
taichinakamura
0
600
生成AIとM5Stack / M5 Japan Tour 2025 Autumn 東京
you
PRO
0
240
Git in Team
kawaguti
PRO
3
330
Where will it converge?
ibknadedeji
0
200
Vibe Coding Year in Review. From Karpathy to Real-World Agents by Niels Rolland, CEO Paatch
vcoisne
0
120
Developer Advocate / Community Managerなるには?
tsho
0
120
綺麗なデータマートをつくろう_データ整備を前向きに考える会 / Let's create clean data mart
brainpadpr
3
370
Featured
See All Featured
The Myth of the Modular Monolith - Day 2 Keynote - Rails World 2024
eileencodes
26
3.1k
Making Projects Easy
brettharned
119
6.4k
Docker and Python
trallard
46
3.6k
The Art of Programming - Codeland 2020
erikaheidi
56
14k
How to Ace a Technical Interview
jacobian
280
24k
Building Adaptive Systems
keathley
43
2.8k
Statistics for Hackers
jakevdp
799
220k
Intergalactic Javascript Robots from Outer Space
tanoku
273
27k
Into the Great Unknown - MozCon
thekraken
40
2.1k
Rails Girls Zürich Keynote
gr2m
95
14k
Templates, Plugins, & Blocks: Oh My! Creating the theme that thinks of everything
marktimemedia
31
2.5k
Performance Is Good for Brains [We Love Speed 2024]
tammyeverts
12
1.2k
Transcript
©2020 Wantedly, Inc. 23th place solution Kaggle Google Quest Q&A
Labeling লձ Feb 28, 2020 - Shuhei Goda - @jy_msc
©2020 Wantedly, Inc. Team - The Hand Shuhei Goda @jy_msc
Visit Engineering Team at Wantedly Naomichi Agata @agatan_ People Engineering Team at Wantedly
©2020 Wantedly, Inc. Model Pipeline #FSUCBTF VODBTFE -JHIU(#. #FSUCBTF VODBTFE
Settings ɾ3fold with GroupKFold ɾBCE + margin ranking loss ɾ3epoch Settings ɾmax_depth=1 ɾlr=0.1 Meta features ɾtext length ɾstackexchange Text data ɾquestion_title ɾquestion_body ɾanswer 1SF1SPDFTT 2BOE" 1SF1SPDFTT POMZ2 ɾquestion_title ɾquestion_body ɾquestion_title ɾquestion_body ɾanswer Settings ɾhtml escape ɾhead+tail truncation
©2020 Wantedly, Inc. ɾHTMLจࣈྻͷΞϯΤεέʔϓ Pre-Process IUUQTXXXLBHHMFDPNDHPPHMFRVFTUDIBMMFOHFEJTDVTTJPO
©2020 Wantedly, Inc. ɾςΩετσʔλͷ݁߹ͱτϦϛϯά ɹɾ[CLS] + question_title + [SEP] +
question_body + [SEP] + answer ɾquestion_body ͱ answer ͕ࢦఆͷ͞Λ͑ͨ߹, ͔྆ΒಉαΠζΛτϦϛϯά Pre-Process IUUQTBSYJWPSHBCT
©2020 Wantedly, Inc. ɾBert-base (uncased) ɹɾޙΖ4ͭͷӅΕͷग़ྗΛ༻ https://arxiv.org/abs/1905.05583 ɹɾQAؒͷSEP tokenͷग़ྗΛ༻ Model
Architecture
©2020 Wantedly, Inc. ɾLabel weight ɹɾ؆୯ͦ͏ͳλεΫweightΛখ͘͞, ෆۉߧͰͦ͠͏ͳλεΫweightΛେ͖͘ ɹɾgpyoptͰweightͷ୳ࡧΛࢼͨ͠Έ͕ͨ, Լهͷ୯७ͳΓํ͕࠷ྑ͔ͬͨ Loss
function Label weight ͋Γ Public: 0.45979, Private: 0.41440 Label weight ͳ͠ Public: 0.43455, Private: 0.40602
©2020 Wantedly, Inc. ɾBCE + margin ranking loss (1 :
1) ɹɾϛχόονΛ2ͭʹׂͯ͠ margin ranking loss Λܭࢉ Loss function BCE + margin ranking loss Public: 0.45979, Private: 0.41440 BCE Public: 0.44006, Private: 0.40668
©2020 Wantedly, Inc. ɾQuestion Model ɹɾQ༻ͷλεΫΛQuestion text͚ͩΛͬͯղ͘ ɹɾΠϯϓοτQ͚ͩͰ͍͍ͷͰ, Qͷtruncationͷྔ͕ݮΔ (Qͷใྔ͕૿͑Δ)
Training Q model + Q and A model Public: 0.45979, Private: 0.41440 Q and A model × 2 (seed average) Public: 0.44298, Private: 0.40613
©2020 Wantedly, Inc. ɾLightGBM ɹɾmax_depth=1, lr=0.1 ɹɾmeta features ɹɹɾtext length
(question, answer) ɹɹɾmeta data from stackexchange (Score, View, FavoriteCount, …) Post-Process LightGBM Public: 0.45979, Private: 0.41440 Simple binning without meta features Public: 0.45282, Private: 0.41387
©2020 Wantedly, Inc. Why we used LightGBM? 1. Simple binning
method ɹɾ༧ଌΛࢄԽ͢Δ͜ͱͰ Spearman’s correlation ͕ྑ͘ͳΔ͜ͱʹؾͮ͘ ɹɾtarget͝ͱʹϏϯαΠζΛࣄલʹઃఆͯ͠Ϗϯೋϯά ɹɾϏϯαΠζݻఆʹ্ͨ͠ͰBertͷ֤epochͷग़ྗΛweighted average (weight࠷దԽ)
©2020 Wantedly, Inc. Why we used LightGBM? 2. Optimize bin-size
and weights ɹɾϏϯαΠζ࠷దͳΛ͍ͨ͘ͳͬͨ ɹɾϏϯαΠζͱweightsͷಉ࣌࠷దԽ্͕ͨ͠ख͍͔͘ͳ͍ ɹɾ࠷దͳϏϯαΠζ༧ଌͷܗʹΑܾͬͯ·Δ. ֤foldͷ࠷దͳϏϯαΠζͷฏۉͱ weighted averageޙͷ༧ଌ࠷దͳͷ͔Βဃ͢Δ
©2020 Wantedly, Inc. Why we used LightGBM? 3. LightGBM ɹɾϏϯαΠζͱweightsͷಉ࣌࠷దԽ͍ͨ͠
ɹɾmeta features͍͍ͨ ɹɾGBDTσʔλΛׂׂͯ͠ޙͷྖҬʹ࠷దͳΛׂΓͯΔख๏ ɹɹˠ ઙ͍߹Ϗϯχϯάͱಉ༷ͷࢄԽ͕Ͱ͖ΔΜ͡Όͳ͍͔ max_depth=2 max_depth=8
©2020 Wantedly, Inc. 4. LightGBM (parameter tuning) ɹɾࢄԽ͢Δ΄Ͳscore͕ྑ͘ͳΔͷͰ, ߏΛۃྗγϯϓϧʹ͍ͨ͠ ɹɾtrainσʔλΛׂͯ͠࠷దͳύϥϝʔλΛݟ͚ͭΔ
ɹɾmax_depthΛҰ൪খ͘͞, lrΛۃྗେ͖ͨ͘͠ํ͕score͕ྑ͘ͳͬͨ Why we used LightGBM?
©2020 Wantedly, Inc. ɾsample weightͷઃఆ ɾhostͷ୯ޠΛΠϯϓοτͷઌ಄ྻʹஔ͘ ɾnew tokenͷՃ ɾBert-base casedΛ͏
ɾtexͷίʔυϒϩοΫΛྗٕͰফڈ Didn’t work for us
©2020 Wantedly, Inc. Discussion: https://www.kaggle.com/c/google-quest-challenge/discussion/129904#742302 Kernel: https://www.kaggle.com/shuheigoda/23th-place-solusion Links
©2020 Wantedly, Inc. https://www.wantedly.com/projects/375150 We are hiring !