�ݺ�ߣ

Inferring User Tasks and Needs
Rishabh Mehrotra1, Emine Yilmaz2, Ahmed Hassan Awadallah3
1Spotify, London
2University College London
3Microsoft Research

Outline of the Tutorial
• Section 1: Introduction
• Section 2: Characterizing Tasks
• Section 3: Tasks Extraction Algorithms
• Section 4: Task based Evaluation
• Section 5: Applications

Section 4: Task Based Evaluation
• User behavior signals
• Predictive Models of SAT
• Explicit Satisfaction Signals

What should we measure?
• From Queries to Tasks
– People do not come to search engines to submit queries, they come
to accomplish tasks
‘‘ If you cannot measure it, you cannot improve it.’’
Lord Kelvin
‘‘ You get what you measure.’’

User Behavior
Signals
Predictive
Models
Explicit Satisfaction
Signals

User Behavior
• Behavioral logs are traces of human behavior seen through the
lenses of a sensor
• In Web search:
• Queries, Clicks, mouse movements, etc.
• In mobile and other devices:
• Voice (acoustics)
• Attention (viewport)

Change in Acoustics
• Slower speech rate is more prevalent
when ASR quality is bad
• Loudness is the perception of the
strength or weakness of a sound
wave resulting from the amount of
pressure produced
• Pitch represents how high or low a
sound is perceived by the human ear
and is determined by a
sound's frequency
0%
5%
10%
15%
20%
25%
30%
35%
40%
0% 10% 20% 30% 40% 50% 60% 70% 80% 90%100%
% requests
slower ratio r
SAT ASR Quality
DSAT ASR Quality
[Kulkarni et al., ICASSP 2017]

Attention Modelling
• Viewport is the portion of the page
that is visible on the screen
• There is a high correlation between
gaze time and viewport time on
Mobile devices
[Lagun et al., SIGIR’14]

Tasks as a Trail
ENDtataSTARTG nn ,,,.......,, 11=
Goal 1: Q 4s RL 1s SR 53s SR 118s END
Goal 2: Q 3s Q 5s SR 10s AD 44s END
Goal 3: Q 4s RL 1s SR 53s SR 118s END
• A user search task can be represented by:
• An ordered sequence of actions
• Time between actions

Modeling action sequences: Markov Model
• Learn patterns of action sequences
that lead to satisfaction/dissatisfaction
• A mixture model for generating
behavior trails with two mixture
components corresponding to
satisfaction and dissatisfaction
[Hassan et al., WSDM 2010]

• Accuracy: Much
better than baselines
on labeled data
• Sensitivity: Much better
than existing metrics for
A/B testing
[Hassan et al., WSDM 2010]

Modeling action sequences: CRF
[Ageev et al., SIGIR 2011]

Semi-supervised Model
• Can we learn from both labeled and unlabeled data?
– Labeled data is typically limited
– Unlabeled data is available at a larger scale
• Generative Model + EM
– E Step: Use the current classifier
to estimate class probabilities
for unlabeled data
– M Step: Re-estimate model
parameters using the labeled data
and the component membership of
the unlabeled data
Model
Expectation Step
Maximization Step
[Hassan, SIGIR 2012]

• Learning from both labeled and unlabeled data
significantly improves the performance
[Hassan, SIGIR 2012]
Semi-supervised Model

Personalized Model
• There are large differences between
users
• A one-size-fits-all model of user
behavior cannot capture the variance in
behavior associated with satisfaction
• Making generalizations about particular
behaviors is risky
[Hassan and White, CIKM 2013]

Dialog Models: From Search to Intelligent Assistants
• In intelligent assistants, we
have a dialog between the user
and the system
• Classify every action/response
to one of a set of predefined
types
• Extend the model to cover
multiple user/system responses
User Request System Response
“Where is the nearest pharmacy” “Here are 8 pharmacies near you.” [show options on
the screen]
“Send me the directions to block
sponsee”
(Show me the directions to Clark’s
pharmacy)
“Sorry, I couldn’t find anything for ‘Send me the
directions to block sponsee.’ Do you wanna search
the web for it?”
“No” “Here are 8 pharmacies near you.” [show options on
the screen]
“Directions to Clark’s pharmacy” “OK, getting you directions to Clark’s Pharmacy.”
[navigation]
[Jiang et al., WWW 2015]

• Joint training of unified Bi-LSTMs & CNN
• Interaction layer: between components of intermediate representation
• Softmax layer at the end for prediction
Unified Multi-View Model
[Mehrotra et al., CIKM 2017]

• Adding the auxiliary SERP level features help
• Proposed unified model performs best across the board
Deep multi-view model performs better than traditional sequential
models!
[Mehrotra et al., CIKM 2017]
Unified Multi-View Model

Learn from the users
• We need models that can predict user
satisfaction using implicit signals from
user interactions
• Find correlation between implicit
behavior signals and some explicit
satisfaction signal

Explicit Satisfaction Signals
• Using Judges/Annotators
– Recreate the user experience and ask an annotator to assess user
satisfaction
[Jiang et al., WSDM 2015]

• Lab Studies
– A group of people who are asked to perform certain tasks and can be
subsequently interviewed to get feedback on satisfaction
– Gamification

• Field Studies
– Richer Client Instrumentation (e.g. Toolbars)
– Users install a special software that monitors their tasks and
collects feedback from them at specific points
– Example: Curious Browser, Search TrailBlazer, SearchVote, etc.
[Fox et al., TOIS 2005; Hassan et al., CIKM 2011]

• Data gathering A/B tests
– Run an A/B where the control group is subjected to a degraded
experience
50%
Users
50%
Users
[Machmouchi et al., CIKM 2017]

Summary: Section IV
• User behavior signals
– Acoustics for voice interactions
– Attention modeling in viewport
– Action sequences
• Predictive Models of SAT
– Markov model for action sequences
– CRF models
– Semi-supervised model
– Deep sequential model for task SAT
• Explicit Satisfaction Signals
– Judges/annotators
– Lab studies
– Field studies
– Data gathering A/B tests

�ݺ�ߣ

Part 4: WWW 2018 tutorial on Understanding User Needs & Tasks

Recommended

More Related Content

Similar to Part 4: WWW 2018 tutorial on Understanding User Needs & Tasks (20)

Recently uploaded (20)

Part 4: WWW 2018 tutorial on Understanding User Needs & Tasks