Ming Sun Yun-Nung Chen
Alexander I. Rudnicky School of Computer Science
Carnegie Mellon University
An Intelligent Assistant for High-Level Task
Understanding
Many devices, many apps…
2
Rich functionality is now available
How well can it
support user
activities?
Book a restaurant!
I can do Yelp!
4
Show me bus to China Palace.
I know Maps!
Tell kids we will have Chinese tonight.
Multi-domain Dialog – Current System
System: What can I do for you?
User: Could you arrange a dinner for me and my friends?
System: Sorry I don’t understand that. What can I do for you?
User: Can I book a table for three in Täkō downtown for this Friday?
…
System: OK. What can I do for you next?
User: Show me the bus from here.
System: Where is your destination please?
User: Täkō downtown Pittsburgh.
…
System: What should I do next?
User: Send the bus route to Carrie and Peter.
System: OK. Sending the following message to Carrie and Peter: “the bus route”.
6
Cannot handle complex intention
Passively support cross-domain dialog
No shared context
Multi-domain Dialog – Human Assistant
Assistant: What can I do for you?
User: Could you arrange a dinner for me and my friends?
Assistant : What kind of food do you prefer?
User: Mexican?
Assistant : How about Täkō? I can book a table for you.
User: Sounds good! Can I take a bus there?
Assistant : 61 A/B/C/D can take you there. Do you want to send this to your friends?
User: Great! Send it to Carrie and Peter.
Assistant: OK. The bus route 61 has been sent.
Understand complex intentions
Actively support cross-domain dialog
Maintain context
Intention Understanding
Yes Dad!
Special Team 1!
8
Plan a dinner!
Intention Understanding
Find a restaurant!
Yelp!
Intention Understanding
Find a bus route!
10
Maps!
Intention Understanding
Message Agnes & Edith!
Messenger!
Intention Understanding
Yes Dad!
Special Team 2!
12
Set up meeting!
Approach
Step 1: Observe human user perform multi-domain tasks
Step 2: Learn to assist at task level
Map an activity description to a set of domain apps
Interact at the task level
Data Collection 1 – Smart Phone
Wednesday 17:08 – 17:14
CMU Messenger Gmail Browser Calendar
Schedule a visit to CMU Lab
Log app invocation + time/date/location
Separate log into episodes if there is 3 minute inactivity
14
Data Collection 2 – Wizard-of-Oz
Find me an Indian place near CMU. Yuva India is nearby.
Monday 10:08 – 10:15
Home Yelp Maps Messenger
Schedule a lunch with David.
Music
Data Collection 2 – Wizard-of-Oz
16
When is the next bus to school? In 10 min, 61C.
Monday 10:08 – 10:15
Home Yelp Maps Messenger
Schedule a lunch with David.
Music
Data Collection 2 – Wizard-of-Oz
Tell David to meet me there in 15 min. Message sent.
Monday 10:08 – 10:15
Home Yelp Maps Messenger
Schedule a lunch with David.
Music
Corpus
533 real-life multi-domain interactions from 14 real users
12 native English speakers (2 non-)
4 males & 10 females
Mean age: 31
Total # unique apps: 130 (Mean = 19/user)
18
Resources Examples Usage
App sequences Yelp->Maps->Messenger structure/arrangement Task descriptions “Schedule a lunch with David” nature of the intention,
language reference User utterances “Find me an Indian place near CMU.” language reference Meta data Monday, 10:08 – 10:15, Home contexts of the tasks
Intention Realization
Model
• [Yelp, Maps, Uber]
• November, weekday, afternoon, office
• “Try to arrange evening out”
• [United Airlines, AirBnB, Calendar]
• September, weekend, morning, home
• “Planning a trip to California”
• [TripAdvisor, United Airlines]
• July, weekend, morning, home
• “I was planning a trip to Oregon”
“Plan a weekend in Virginia”
• […, …, …]
• …
•“Shared picture to Alexis”
Find similar past experience
Cluster-based:
K-means clustering on user generated language
Neighbor-based:
KNN
1
2
3
Cluster-based Neighbor-based
20
Yelp Yelp
OpenTable Yelp
Maps Maps Maps Maps
Messenger Email
Email Email
Realize domains from past experience
Representative Sequence
Multi-label Classification
App sequences of similar experience (ROVER)
Some Obstacles to Remove
Language-mismatch
Solution: Query Enrichment (QryEnr)
[“shoot”, “photo”] -> [“shoot”, “take”, “photo”, “picture”]
word2vec, GoogleNews model
App-mismatch
Solution: App Similarity (AppSim)
Functionality space (derived from app descriptions) to identify apps
Data-driven: doc2vec on app store texts
Rule-based: app package name
Knowledge-driven: Google Play similar app suggestions
22
Gap between Generic and Personalized Models
QryEnr, AppSim, QryEnr+AppSim reduce the gap of F1
10 15 20 25 30 35
Compare different AppSim
0 10 20 30 40 50 60 70
Precision Recall F1
Baseline Data Knowledge Rule Combine
24
Compare different AppSim
Combining three approaches performs the best
Knowledge-driven and data-driven have low coverage among (manufacture) apps
Rule-based is better than the other two individual approaches
Learning to talk at the task level
Techniques:
(Extractive/abstractive) summarization
Key phrase extraction [RAKE]
User study:
Key phrase extraction + user generated language
Ranked list of key phrases + user’s binary judgment
1. solutions online 2. project file
3. Google drive 4. math problems
5. physics homework 6. answers online
…
[descriptions]
Looking up math problems.
Now open a browser.
Go to slader.com.
Doing physics homework.
…
[utterances]
Go to my Google drive.
Look up kinematic equations.
Now open my calculator so I can plug in numbers.
…
26
1. solutions online 2. project file
3. Google drive 4. math problems
Learning to talk at the task level
Metrics
Mean Reciprocal Rank (MRR)
Result:
MRR ~0.6
understandable verbal reference show up in top 2 of the ranked list
[descriptions]
Looking up math problems.
Now open a browser.
Go to slader.com.
Doing physics homework.
Summary
Collected real-life cross-domain interactions from real users
HELPR: a framework to learn assistance at the task level
Suggest a set of supportive domains to accomplish the task
Personalized model > Generic model
The gap can be reduced by QryEnr + AppSim
Generate language reference to communicate verbally at task level
28
HELPR demo
Interface
HELPR display
GoogleASR
Android TTS
HELPR server
User models
Thank you
Questions?
30
