Transcript We Feel Fine and Searching the Emotional Web Sepandar D

We Feel Fine and
Searching the Emotional Web
Sepandar D. Kamvar, Jonathan Harris
Stanford University, Number 27
WSDM ’11
06 April, 2011
Hye Chan, Bae
Outline







Introduction
Design Considerations
Architecture
User Interface
API
Discussion
Conclusion
2/36
Introduction
 Sentiment analysis
– The growth of the social web has led to an increased its interest in
Sentiment analysis
3/36
Introduction
 Sentiment analysis
– Typical applications have helped consumers make purchase decisions
 E.g. “thumbs up” / “thumbs down”
happy
4/36
Introduction
 Sentiment analysis
– The large-scale availability of emotional text gives the ability to better
understand emotions themselves
5/36
Introduction
 We Feel Fine
– A project that aims to collect the world’s emotions
(since August 2005)
– Searches the phrases “I feel” and “I am feeling”
– Identifies the “feeling” and extracts a number of demographic information
– Using a series of playful interfaces, the feelings can be searched offering
responses to specific questions
6/36
Outline







Introduction
Design Considerations
Architecture
User Interface
API
Discussion
Conclusion
7/36
Design Considerations
 Sentence-level analysis
– People often express emotions at the sentence level; rarely is an entire
document about a single emotion
 Indexing context
– There is much useful context to an emotion outside of the words
(time, location, gender, age of the person)
How do women feel right now?
How did people in the U.S. feel on September 11th?
8/36
Design Considerations
 Sentiment as the primary organizing principle
– The primary aim is to understand more about emotions themselves
 De-emphasizing ranking
– It is much more difficult to rank sentiment
– Thousands of different expressions can be equally reasonable responses
– No ranking in We Feel Fine
“feelings are never wrong”
9/36
Design Considerations
 Emphasizing browsing and summarization
– Users can gain intuition through qualitative exploration
– Allowing the user to quickly get the gestalt of how a population feels
 Enabling the user to easily shift between macro and micro
– Macro-level (summarization)
– Micro-level (browsing)
Micro
Macro
10/36
Design Considerations
 Visualizations that reflect the data
– An ideal UI should reflect the subject matter
 Direct Access to the Data
– For both an artwork and a scientific tool, it provides a data API for direct
data access
11/36
Outline







Introduction
Design Considerations
Architecture
User Interface
API
Discussion
Conclusion
12/36
Architecture
URLServer
List of urls
 blog posts
 microblog feeds
 public social network msg
Crawler
Designed so that can easily
add more crawling machines
Emotional Lexicon
Feeling
Indexer
Weather Server
Image Repository
13/36
Architecture
Feeling
Indexer
Image Repository
Feeling sentences
& metadata
Query Cache
We Feel Fine
Database
 MySQL
 replicated database server
 designed to be easily sharded by date
API Server
defines a RESTful API
Sentiment Mining Server
14/36
Montage
Server
Architecture
Montage
Server
API Server
We Feel Fine
Frontend
Third-Party
Applications
Java applet
15/36
Montage
Gallery
Architecture
16/36
Outline







Introduction
Design Considerations
Architecture
User Interface
API
Discussion
Conclusion
17/36
User Interface
 Search Panel
– Allows the view to choose the sample population
– Can select any combination of the following axes
 Feeling, Age, Gender, Weather, Location, Date
18/36
User Interface
 Madness
– A playful interface to interact with individual data items
– Each particle represents a single feeling
19/36
User Interface
 Murmurs
– Presents a structured environment in which to view feelings
20/36
User Interface
 Montage
– Presents the feeling from a given population that contain photographs
– Any user can save a montage to the Montage Gallery
 Allowing anonymous viewers to curate an exhibit of interesting images
21/36
User Interface
 Mobs
– Consists of five smaller movements
 feeling, gender, age, weather, location
– Aims to summarize the data set
22/36
User Interface
 Metrics
– Also consists of five smaller movements
– Expresses the features that are most differentially expressed from the
global average
23/36
User Interface
 Mounds
– Displays every feeling in database
– Each feeling is portrayed as a large bulbous mound
24/36
User Interface
 Usage Observation
– Emotional Self-Awareness
 the subject started talking about how she felt around the middle or end of the
session
 Many participants also noticed that their own emotions mirrored those of the
people in the piece
– Empathy
 Participants reported a feeling of connection and empathy
 They project their own experience on to the emotions they see
25/36
Outline







Introduction
Design Considerations
Architecture
User Interface
API
Discussion
Conclusion
26/36
API
 2 components
– RESTful API
 Translates a url to a SQL query on database
 Returns the results in XML, HTML, CSV or plain text
 User can query by some conditions
– Sentiment Mining Server
 A set of functions that postprocess an APU query to compute statistics
– Frequency histogram, breakdown, categorize feelings, etc.
 Support a wide array of uses
– Has been accurate both in psychology literature and in new hypotheses
27/36
API
 Usage Observation
– The Meaning of Happiness
 The co-occurrence of excited and happy feelings for younger people
 The co-occurrence of peaceful and happy feelings for older people
– Hedonometer
 it has been built based on We Feel Fine data and the ANEW scoring system
– The Emotions of Aging
 People’s emotions vary with age
28/36
API
29/36
API
30/36
API
 Usage Observation
– Time-series Analyses
31/36
API
 Usage Observation
– The Emotional Graph
 Shows emotions that are frequently co-expressed in the same sentence
32/36
API
 Usage Observation
– Artistic Purposes
 Prayer Companion
 An installation in Denmark city hall tower
 A robot that mixes a drink based on the feelings returned by We Feel Fine
33/36
Outline







Introduction
Design Considerations
Architecture
User Interface
API
Discussion
Conclusion
34/36
Discussion
 Unintended and broad-reaching consequences
– Experiential Data Visualization
 The primary responses in the user study were not cognition but affective
 3 properties of EDV
– Communicate insights that are often simply communicated in words
but much more powerfully communicated by example
(love are easily expressed in words but more powerfully expressed by being in love)
– Focus on interaction models that encourage direct interaction with individual data
items
– Focus on influencing affect rather than cognition
 The design principles in section2 are useful guiding principles for EDV in general
35/36
Discussion
 Unintended and broad-reaching consequences
– Crowdsourced Data Mining
 Potential of crowdsourced data analysis
– Over 8 million people spent an avg of 4 minutes exploring the data
– Equivalent to a staff of over 50 people working full-time
 Unique about We Feel Fine
– Include not only statistics but detailed examples
(crowdsourced qualitative research)
 Aggregating, communicating, corroborating the insights of the crowd more
seamlessly is an area of future work
36/36
Outline







Introduction
Design Considerations
Architecture
User Interface
API
Discussion
Conclusion
37/36
Conclusion
 Item-level of exploration of data in immersive interface
– bring experiential benefits
– enable crowdsourced qualitative data analysis
 Can be used to be tools to support social science research
– Allows to run inexpensive large-scale studies to generate data-driven
hypotheses
38/36
Thank you!!