kantanfest: dimitar shterionov - part 2
TRANSCRIPT
KantanNeural™ from A to Z3/3: NMT in 4 weeks → 4 days → 4 hours
Dimitar Shterionov
What is NMT?
31/07/2017 KantanFest, Dublin, Ireland 2
What is NMT?
31/07/2017 KantanFest, Dublin, Ireland 3
x1 x2 x3 c y1 y2 y3
How to NMT – The Recipe
Hardware + Software: GPUs, torch, Theano
nematus, OpenNMT
Know-how, Support
Integration,Deployment
Training data
31/07/2017 KantanFest, Dublin, Ireland 4
How to NMT – KantanNeural™
Hardware + Software: GPUs, torch, theano
nematus, OpenNMT
Know-how, Support
Integration,Deployment
Training data
KantanNeural™
31/07/2017 KantanFest, Dublin, Ireland 5
KantanNeural™: black board to production
Proof of Concept:
AWS, NVIDIA K520 GPUs
Nematus, ADAM, BPE, SCN
MT (engines) build: 4 weeks
Quality: impressive
01 Nov 2016
31/07/2017 KantanFest, Dublin, Ireland 6
• ADAM: Parameter update algorithm
• Byte-pair encoding (BPE)• Single-character n-gram (SCN)
lower → low ertallest → tall estalmost → al most
lowesttallerallow
KantanNeural™ α:
OpenNMT, ADAM, BPE
ΜΤ build time: 4 days
Quality: on a par with nematus
KantanFleet™
01 Nov 2016 01 Feb 2017
KantanNeural™: black board to production
31/07/2017 KantanFest, Dublin, Ireland 7
KantanNeural™ β:
Build-your-own NMT
Available to all clients (no extra charge)
Extended KantanFleet™
01 Nov 2016 01 Feb 2017 15 March 2017
KantanNeural™: black board to production
31/07/2017 KantanFest, Dublin, Ireland 8
01 Nov 2016 01 Feb 2017 15 March 2017
Currently:
Build-your-own NMT
NVIDIA K80 GPUs
AdaptiveMT
Incremental Retraining
4 hours?
30 June 2017
31/07/2017 KantanFest, Dublin, Ireland 9
KantanNeural™: black board to production
KantanMT.com – A Complete Platform
Build
Improve
Deploy
Kantan
Templates
Kantan
NER
Kantan
Llibrary
Kantan
Fleet
Kantan
BuildAnalytics
Kantan
Analytics
Kantan
PEX
Kantan
LQR
Adaptive
MT
Kantan
GENTRY
Kantan
TotalRecall
KantanNeural™Kantan
Translate
Kantan
Swift
Kantan
API
Kantan
AutoScale
Kantan
OfficeMT
Kantan
Connectors
Kantan
Snippets
KantanNeural™
31/07/2017 KantanFest, Dublin, Ireland 10
KantanMT.com – A Complete Platform
Build Improve Deploy
31/07/2017 KantanFest, Dublin, Ireland 11
KantanMT.com – A Complete Platform
Build Improve Deploy
Select a KantanFleet™ engine
KantanFleet™ Neural (18 language
pairs)
Multiple domains
Create new NMT engine
Import library data
Import your own data
Convert an SMT profile:
… just two clicks away from NMT
31/07/2017 KantanFest, Dublin, Ireland 12
KantanMT.com – A Complete Platform
Build Improve Deploy
Select a KantanFleet™ engine
31/07/2017 KantanFest, Dublin, Ireland 13
KantanMT.com – A Complete Platform
Build Improve Deploy
Create a blank KantanNeural™
engine
31/07/2017 KantanFest, Dublin, Ireland 14
KantanMT.com – A Complete Platform
Build Improve Deploy
Convert a PBSMT engine into KantanNeural™
engine
31/07/2017 KantanFest, Dublin, Ireland 15
KantanMT.com – A Complete Platform
Build Improve Deploy
31/07/2017 KantanFest, Dublin, Ireland 16
Artificial Neural Networks train iteratively:
While stopping condition not met:
While training data not exhausted:
Take a batch
Learn from it
Repeat
KantanMT.com – A Complete Platform
Build Improve Deploy
Augment data
Parallel corporaPreprocessing rules
(PEX, tokeniser excep., etc.)
F-Measure, BLEU, TERKantanLQR(Error typology, AB Testing)
New Preprocessing rulesNew data
Augment data Augment data Augment data
31/07/2017 KantanFest, Dublin, Ireland 17
KantanMT.com – A Complete Platform
Build Improve Deploy
Augment data
Parallel corporaPreprocessing rules
(PEX, tokeniser excep., etc.)
F-Measure, BLEU, TERKantanLQR(Error typology, AB Testing)
New Preprocessing rulesNew data
Augment data Augment data Augment data
31/07/2017 KantanFest, Dublin, Ireland 18
KantanMT.com – A Complete Platform
Build Improve Deploy
Augment data
Parallel corporaPreprocessing rules
(PEX, tokeniser excep., etc.)
F-Measure, BLEU, TERKantanLQR(Error typology, AB Testing)
New Preprocessing rulesNew data
Augment data Augment data Augment data
31/07/2017 KantanFest, Dublin, Ireland 19
4 hours?
KantanMT.com – A Complete Platform
Build Improve Deploy
API
Connectors
KantanWidgets™
As every other KantanMT engine
31/07/2017 KantanFest, Dublin, Ireland 20
Conclusions…
KantanMT:
A complete MT platform for both NMT and PBSMT engines
Easy access to powerful MT technology
How to train, improve and deploy KantanNeural™ engines
Seamless switch from PBSMT to NMT
Incremental retraining to improve, adapt and specialize engines
Conclusions…
KantanMT:
A complete MT platform for both NMT and PBSMT engines
Easy access to powerful MT technology
How to train, improve and deploy KantanNeural™ engines
Seamless switch from PBSMT to NMT
Incremental retraining to improve, adapt and specialize engines
4 hours training?
… and future work
Better control:
Terminology
Tags
NTAs
Learn from postedits:
Exploit feedback from KantanLQR™
Exploit feedback from connectors
Models:
Add language knowledge
Hybrid MT
Convolutional Neural Networks (CNN)
…
Solving
Thank you…
Laura Casanellas: [email protected] Shterionov: [email protected]
KantanLabs: [email protected]
KantanMT: [email protected]