Search Text Mining Web Site Usability

1
SearchText MiningWeb Site Usability

• Marti Hearst
• SIMS

2
BAILANDO Projects

• Better Access to Information
• using Language Analysis and
• Novel Dynamic Organizations

3
Current BAILANDO Projects

• CHA-CHA FLAMENCO
• Better Search Interfaces
• LINDI
• UI support for Search
• Text Data Mining
• TANGO
• Automated Web Site Usability

4
Search UIs

• Combine Browsing Search
• Place Search Results in Context

Large Category Hierarchies
5
Cha-Cha Students Mike Chen, Jamie Laflen, Jason
Hong, Jimmy Lin, Shiang Chen
6
Medical Category Hierarchy
7
DynaCat (Pratt, Hearst, Fagan 99)
8
DynaCat Study

• Design
• Three queries
• 24 cancer patients
• Compared three interfaces
• ranked list, clusters, categories
• Results
• Participants strongly preferred categories
• Participants found more answers using categories
• Participants took same amount of time with all
  three interfaces
• Similar results have been verified by another
  study by Chen and Dumais (CHI 2000)

9
Cat-a-Cone Interface(Hearst Karadi 97)
10
FLAMENCOImproving Search via Large Category
Hierarchies

• How to show intersections across category types?
• How to preview related categories in a
  user-tailored, dynamic manner?

11
Text Data Mining

• Relationships between information in documents
  can create new facts, not previously known.

12
Imagine

• You are a medical researcher
• Your patient has
• spinal inflammation
• numbness in fingers
• low TC levels
• negative results for all tests
• How can you help her?

13
Idea

• A new way of searching text.
• Link pieces of information together
• to formulate hypotheses

14
LINDILinking Information for New DIscoveries

• Three main parts
• Search UI for building and reusing hypothesis
  seeking strategies.
• Statistical language analysis techniques for
  interpreting the text.
• Backend for interfacing with various databases
  and translating different formats.

15
Gathering Evidence
Spinal Inflammation
Numbness in fingers
Low TC Levels
16
Gathering Evidence
Find diseases associated with each
Spinal Inflammation
Numbness in fingers
Low TC Levels
17
Gathering Evidence
Find unanticipated commonalities
Spinal Inflammation
Numbness in fingers
Low TC Levels
18
Supporting Cascaded Search Operations
19
(No Transcript)
20
New Language Analysis

• First use category labels to retrieve candidate
  documents
• Then use language analysis to detect causal
  relationships between concepts
• Title
• Magnesum deficiency implicated in increased
  stress levels.
• Interpretation
• ltnutrientgtltreductiongt related-to
  ltincreasegtltsymptomgt
• Use these to find relationships and formulate
  hypotheses

21
Statistical Semantic Parsing

• Modern statistical techniques
• Mainly applied to syntactic structure
• Probabilistic knowledge representation
• Represent hypotheses with different degrees of
  certainty.

22
Automating Assessment of Web Site Usability

23
Why Worry?

• Problem IBM's extranet
• Heavy use of help and search
• Unhappy users
• Solution
• Massive web site redesign
• Focus on info-organization, not the purchasing
  process.
• Cost "in the millions"
• Results
• Not announced or trumped up
• Use of "help" decreased 84
• Sales increased 400

24
Web TANGOTool for Assessing NaviGation
Organization

• Goal automated support for comparing design
  alternatives
• How Assess usability of the information
  architecture
• Approximate peoples information-seeking behavior
  (Monte Carlo simulation)
• Output quantitative usability metrics

25
Guidelines

• There are many usability guidelines
• A survey of 21 sets of web guidelines found
  little overlap (Ratner et al. 96)
• Why?
• Our hypothesis not empirically validated
• So lets figure out what works!

26
An Empirical Study
Which features distinguish well-designed web
pages?
27
Methodology

• Data collection
• 1108 pages
• 163 sites
• 3 levels per site
• 14 metrics
• About 85 accurate
• Text cluster and text positioning counts less
  accurate

28
Metrics
29
Preliminary Results

• Linear regression to predict Webby judges ratings
• Top 30 vs bottom 30
• Prediction accuracy
• 72 if categories not taken into account
• 83 if categories assessed separately

30
Goals

• Create empirical foundations for what is still
  guesswork
• Next step
• A free online tool
• Long term goal
• An monte carlo simulator for comparing potential
  designs

31
For More Information

• http//webtango.berkeley.edu
• hearst_at_sims.berkeley.edu