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
Google_QA__Solution___Review__1_.pdf
Search
Draconda
February 28, 2020
3
2.6k
Google_QA__Solution___Review__1_.pdf
Draconda
February 28, 2020
Tweet
Share
Featured
See All Featured
XXLCSS - How to scale CSS and keep your sanity
sugarenia
240
1.2M
Making Projects Easy
brettharned
108
5.5k
Fireside Chat
paigeccino
20
2.6k
Adopting Sorbet at Scale
ufuk
67
8.6k
Web Components: a chance to create the future
zenorocha
305
41k
No one is an island. Learnings from fostering a developers community.
thoeni
14
2.1k
How to Create Impact in a Changing Tech Landscape [PerfNow 2023]
tammyeverts
12
1.5k
JavaScript: Past, Present, and Future - NDC Porto 2020
reverentgeek
39
4.4k
Exploring the Power of Turbo Streams & Action Cable | RailsConf2023
kevinliebholz
1
3.4k
GitHub's CSS Performance
jonrohan
1023
450k
Building Adaptive Systems
keathley
30
1.8k
Faster Mobile Websites
deanohume
297
30k
Transcript
Google QA Solution & Review
Agenda ・Model Architecture ・Post Processing ・Augmentation ・Other Tips & Trials
Model Architecture ・3 models in 1 model ・MLE Layer ・Ensemble
of Bert/XLNet
Post Process ・Threshold Optimization ・Test N for range(200) & Select
Best CV
Augmentation ・Back Translation ・TextBlob
+1. Flex Module ・Change MaxLen for Token selection Max Len
== Max Len1 + Max Len2 == 512 - 3
+1. MLE Module ・Almost Same ・Ensemble is also well VS
+α. Bert Pretraining (MLM) ・Training on External Dataset
+α. Distillation ・Pseudo label for External Dataset
+α. Other Models ・Bert-Large, XLNet-Large ・RoBerta, RoBerta-Large ・ALBert, ALBert-Large ・GPT2,
XLM, etc...
+α. Pseudo Labeling ・1 Opinion about PSeudo Labeling
+α. Draco PP V2 ・For 12 columns which (Max -
Min) < 1 ・Invert when evaluate
+α. Post Process+1 ・for N in range(200) Optimize R ・Scipy,
NN, etc...
Thanks !!