WeChat - Nvidia

WeChat
Social to Intelligent Connection
杨强
What is WeChat?
A New Lifestyle
• Mobile App for 500M+
• Multimedia
• A Way to Connect: A Platform
Users(Unit:100M)
12
>10
Rapid Growth
10
1 Billion Accounts, 550M Active Users
8
6
8M Service Accts
4
20 Languages, 200 Countries
2
0
Oversea Users
By Aug 2013
Active Users
per Month
By Sep 2014
1.0
4.68
微信 Growing Up Path
~6
Moment Ads
Growth Path (Unit: 100M)
2014-10-1
Video Clips
1
Red Packet
3
2
2013-1-15
Payment
Online Shopping
2012-9-17
2012-3-29
Video Chat
Birth
Speech
Nearby
Shake It!
Scan
Service
Platform
2015
-广告
2014
-小视频
2014-10-1
-卡券
3 Milestones
-红包
-支付
2012
2013
-游戏
-表情商店
-视频聊天
3亿
2013-1-15
2亿
2012-9-17
1亿
2012-3-29
-朋友圈
-公众平台
2011
-扫一扫
-摇一摇
-漂流瓶
-附近的人
2010
-语音信息
微信诞生
433 days: from 0 to 100 M
6 Months Later: doubled to 200M
3 months later: 300 M
COLLECTIVE USER EXPERIENCE: SHAKING
FOR RED POCKET IN 2015 SPRING FESTIVAL
•
INTERACTION / COMMENTS / VOTING / LOTTERY / SHARE WITH FRIENDS
Red envelope data:Over 1.01 billion times、Peak at 550 000 / min、Shaking 11
billion times, Peak 810 million times / min
LINK
BigData @ WeChat
WeChat Big Data
contacts
Big Data
relational data
subscribed public
accounts
group
non-relational
behavior
text
explicit
information
posts
images
user feedback
comments
videos
Relationship Types (%)
Friends in Real Life
90
现实生活中的朋友
Classmates
81.4
同学
75.7
亲人或亲戚
70.8
同事
50.4
老师或领导
网友
陌生人
Strangers
32.1
24.7
Artificial Intelligence @
Image Understanding
Lifelong Learning Agent
Spammer and Rumor
detection
Location-based social
networks
NLP and Text
Recommendation
Feature Engineering
Provence of information
Speech Understanding
Social Search
Trust Assessment
User Modeling
Crowd Intelligence
Sentiment Analysis
Event Detection
Network Analysis
WeChat User Modeling &
Transfer Learning
Case Study:
ADs IN WeChat MOMENTS
•
PRESENTS IN MOMENTS / TARGETING BASED ON BIG
DATA / INNTERACTIVE COMMUNICATION
• SPREADING BETWEEN FRIENDS
User Modeling for Ads
• Data Sources
• Demographic information
• Articles read
• Public accounts subscribed
• Techniques:
Source 1
Source 2
Source 3
• DNN
• Multi-task Learning
Source 1
Topic modelling
Source 2
Topic modelling
Source 3
Topic modelling
tag 1
Seed users
Advertise
Co-training
tag n
Image Understanding for Ads
Cross Domain Transfer Learning
• Predicting User Feedback from Social Data
Source domain: BMW advertisement
with user feedback
Target domain: SOHO advertising with
no previous data
Crowd Intelligence @ WeChat
2015
-广告
2014
-小视频
-卡券
Crowd Intelligence
红包
-
2012
2013
-游戏
-表情商店
-视频聊天
3亿
2013-1-15
2亿
2012-9-17
1亿
2012-3-29
-朋友圈
-公众平台
2011
-扫一扫
-摇一摇
-漂流瓶
-附近的人
2010
-语音信息
微信诞生
Grow with Users
- Games
- Red Pocket
- Moments Ads
- Shaking TV Program
Charity by the Millions: Voice Donor
LINK
Crowd Intelligence for the Visually Impaired
Collect Voices
Filter by Standard Models
Speech Recognition
• Large Mandarin Corpus: DNN (deep neutral network)
• Language model:
• N-gram, DNN
• Low-rank matrix
• GPU training
• Decoder:
• WFST framework
• Large, parallel search space
Audio Fingerprinting
• Challenge
• noisy environments,
• compactness of fingerprint, and
• service scalability when song database is
huge (10M)
• Application: WeChat “Shake” Music,
lunched in Jan, 2013
• Big Music Database(10M songs)
Fast Recognition (3-5 seconds)
• Daily Page View > 8M, User View > 3M
Audio Fingerprinting:WeChat Live TV recognition
Recognize live TV program from audio
fingerprinting
• Challenge: High concurrent throughput
• SHAKE-TV:
• Can recognize > 500 TV channels
across China
• User View: 1M simultaneously
• Rich Cross-TV Screen User Experience
• Fully integrated with social networks
Image Understanding @ WeChat
Mariana CNN on GPUs
Mariana CNN is Tencent’s Deep Convolutional Neural Network
based on Single-machine, Multi-core GPU Computation.
• Data parallelism and model parallelism
• Partition models for parallel execution
• Model scalability and performance had major improvement
GPU0
GPU1
GPU2
GPU3
Configuration
Speed-up
2 GPUs Model
Parallelism
1.71
2 GPUs Data Parallelism
1.85
4 GPUs Model + Data
Parallelism
2.52
4 GPUs Data Parallelism
2.67
Mobile Visual Search
• Scan for information or services
• Local and Global Image Feature Descriptor
• Highly Efficient Feature Indexing and Matching
• Mobile video and image quality assessment
• Challenges:
• Variable lighting, Non-planar recognition
• WeChat “Scan” on covers, lunched in 2013
• Large Image databases (~10M)
• Open interfaces for developers
LINK
Mobile Image Tech.
• OCR on a mobile device
• Camera OCR based language
Translation
• Certificate and ID OCR on mobile or
cloud
• Face Technology:
• Detection, Alignment, Tracking,
Recognition/Verification
• User modeling based on images
• Targeted Adverts
LINK
Augmented Reality
3D animation w/ embedded video on
designs
• Rich interaction with users
Challenges:
• Real-time and precise target
detecting/tracking, model rendering
Applications:
• lunched in WeChat movie ticket App “微
票” (Jan 2015)
LINK
Natural Language Understanding
@ WeChat
WeChat NLP
WeChat: Closed-loop NLP
• Closed-loop Feedback in WeChat Services
• Always online: real-time message platform
• Massive user base: 549 million monthly active users
• Payment  User Intention
payment
…
WeChat NLP
Word Multi-Embedding
Learn Embedded Word
Representation
WeChat NLP
Semantic Matching of Questions and Answers
Dependency-Tree RNN model
Semantic Match
WARP loss
• Dependency-Tree RNN
R(query,doc2)
R(query,doc1)
output
•Word multi-embedding match
•BM25
•Other Features:
-Sentence type recognition
-Synonym
-Antonym
-Parsing
output
output
h
h
x
h
x
h
x
query
x
x
h
x
h
x
h
x
h
h
x
doc1
• semantic match
• semantic answer ranking
x
x
x
doc2
NLP & Question Answering
Text/Voice
Speech
Recognition
NLU
intent
Query Analysis
identification
query rewriteQuery Inference Sentiment Analysis
NER/NED Semantic Pattern
Self-learning
Parsing
system
Graph Search
RDF based
Inverted index based
Semantic Search
Data management
Semantic Match
WeChat
knowledge
graph
User Profile
Dialog
Context
Log/Session
User behavior
management
management
management
Text Search
Inverted index
TTS
Answer
NLG
Pattern based
Parsing and LM based
Recommendation
ReRanking
NLP Dialog
微信 Future
Connecting People, Services and Things
Connect Everything
A New Connection Model
People
Things
Becomes a New Lifestyle
Service
Will extend the connection to daily life, commerce &
entertainment
Provide Mobile Internet Service to Industries
People, things and services
An ecosystem for connection, a
new solution provider
Internet +
Intelligent Business Solutions
Integrate internet and other businesses,Smart
Cities,Improved User experience, Connecting
everything
Eight Principles
1、Bring New Value
2、Remove geographical restrictions
3、Remove middle men
4、Distributed
5、Ecosystem
6、Evolutionary Service Platform
7、Social Centric
8、Users’ Interests Always #1
THANK YOU