Using Words to Search a Thousand Images Hierarchical Faceted Metadata in Search

1
Using Words to Search a Thousand
ImagesHierarchical Faceted Metadata in Search
Browsing

• Marti Hearst
• SIMS, UC Berkeley
• Research funded by
• NSF CAREER Grant IIS-9984741

2
Outline

• How do people search for images?
• Current approaches
• Spatial similarity
• Keywords
• Our approach
• Hierarchical Faceted Metadata
• Very careful UI design and testing
• Usability Study
• Conclusions

3
How do people want to search and browse images?

• Ethnographic studies of people who use images
  intensely find
• Find specific objects is easy
• Find images of the Empire State Building
• Browsing is hard, and people want to use rich
  descriptors.

4
Ethnographic Studies

• Garber Grunes 92
• Art directors, art buyers, stock photo
  researchers
• Search for appropriate images is iterative
• After specifying and weighting criteria,
  searchers view retrieved images, then
• Add restrictions
• Change criteria
• Redefine Search
• Concept starts out loosely defined, then becomes
  more refined.

5
Ethnographic Studies

• Markkula Sormunen 00
• Journalists and newspaper editors
• Choosing photos from a digital archive
• Stressed a need for browsing
• Searching for specific objects is trivial
• Photos need to deal with themes, places, types of
  objects, views
• Had access to a powerful interface, but it had 40
  entry forms and was generally hard to use no one
  used it.

6
Query Study

• Armitage Enser 97
• Analyzed 1,749 queries submitted to 7 image and
  film archives
• Classified queries into a 3x4 facet matrix
• Rio Carnivals Geo Location x Kind of Event
• Conclude that users want to search images
  according to combinations of topical categories.

7
Ethnographic Study

• Ame Elliot 02
• Architects
• Common activities
• Use images for inspiration
• Browsing during early stages of design
• Collage making, sketching, pinning up on walls
• This is different than illustrating powerpoint
• Maintain sketchbooks shoeboxes of images
• Young professionals have 500, older 5k
• No formal organization scheme
• None of 10 architects interviewed about their
  image collections used indexes
• Do not like to use computers to find images

8
Current Approaches to Image Search

• Using Visual Content
• Extract color, texture, shape
• QBIC (Flickner et al. 95)
• Blobworld (Carson et al. 99)
• Body Plans (Forsyth Fleck 00)
• Piction images text (Srihari et al. 91 99)
• Two uses
• Show a clustered similarity space
• Show those images similar to a selected one
• Usability studies
• Rodden et al. a series of studies
• Clusters dont work showing textual labels is
  promising.

9
Rodden et al., CHI 2001
10
Rodden et al., CHI 2001
11
Rodden et al., CHI 2001
12
Current Approaches to Image Search

• Keyword based
• WebSeek (Smith and Jain 97)
• Commercial image vendors (Corbis, Getty)
• Commercial web image search systems
• Museum web sites

13
A Disconnect

• Why are image search systems built so
  differently from what people want?
• An image is worth a thousand words.
• But the converse has merit too!

14
Some Challenges

• Users dont like new search interfaces.
• How to show lots more information without
  overwhelming or confusing?

15
Our Approach

• Integrate the search seamlessly into the
  information architecture.
• Use proper HCI methodologies.
• Use faceted metadata
• More flexible than canned hyperlinks
• Less complex than full search
• Help users see where to go next and return to
  what happened previously

16
Faceted Metadata
17
Metadata data about dataFacets orthogonal
categories
18
Faceted Metadata Biomedical MeSH (Medical
Subject Headings)www.nlm.nih.org/mesh
19
Mesh Facets (one level expanded)
20
Questions we are trying to answer

• How many facets are allowable?
• Should facets be mixed and matched?
• How much is too much?
• Should hierarchies be progressively revealed,
  tabbed, some combination?
• How should free-text search be integrated?

21
An Important Trend in Information Architecture
Design

• Generating web pages from databases
• Implications
• Web sites can adapt to user actions
• Web sites can be instrumented

22
A Taxonomy of WebSites
high
Complexity of Data
low
low
high
Complexity of Applications
From The (Short) Araneus Guide to Website
development, by Mecca, et al, Proceedings of
WebDB99, http//www-rocq.inria.fr/cluet/WEBDB/pr
ocwebdb99.html
23
The Flamenco Interface

• Nine hierarchical facets
• Matrix
• SingleTree
• Chess metaphor
• Opening
• Middle game
• End game
• Tightly Integrated Search
• Expand as well as Refine
• Intermediate pages for large categories

24
(No Transcript)
25
(No Transcript)
26
(No Transcript)
27
(No Transcript)
28
(No Transcript)
29
(No Transcript)
30
(No Transcript)
31
(No Transcript)
32
(No Transcript)
33
What is Tricky About This?

• It is easy to do it poorly
• See Yahoo example
• It is hard to be not overwhelming
• Most users prefer simplicity unless complexity
  really makes a difference
• It is hard to make it flow
• Can it feel like browsing the shelves?

34
How NOT to do it

• Yahoo uses faceted metadata poorly in both their
  search results and in their top-level directory
• They combine region other hierarchical facets
  in awkward ways

35
Yahoos use of facets
36
Yahoos use of facets
37
Yahoos use of facets
38
Yahoos use of facets

• Where is Berkeley?
• College and University gt Colleges and
  Universities gtUnited States gt U gt University of
  California gt Campuses gt Berkeley
• U.S. States gt California gt Cities gtBerkeley gt
  Education gt College and University gt Public gt UC
  Berkeley

39
Problem with Metadata Previews as Currently Used

• Hand edited, predefined
• Not tailored to task as it develops
• Not personalized
• Often not systematically integrated with search,
  or within the information architecture in general

40
HCI Methodology

• Identify Target Population
• Needs assessment.
• What to people want how to they work?
• Lo-fi prototyping.
• Produce cheap (throw-away) prototypes
• Get feedback from target population
• Design / Study Round 1.
• Simple interactive version. See if main ideas
  work.
• Design / Study Round 2
• More thorough interactive version more graphics.
  Begin to fine-tune, fix remaining major problems
• Design / Study Round 3
• Continue to fine-tune. Introduce more advanced
  features.

41
Our Project History

• Identify Target Population
• Architects, city planners
• Needs assessment.
• Interviewed architects and conducted contextual
  inquiries.
• Lo-fi prototyping.
• Showed paper prototype to 3 professional
  architects.
• Design / Study Round 1.
• Simple interactive version. Users liked metadata
  idea.
• Design / Study Round 2
• Developed 4 different detailed versions
  evaluated with 11 architects results somewhat
  positive but many problems identified. Matrix
  emerged as a good idea.
• Metadata revision.
• Compressed and simplified the metadata
  hierarchies

42
Our Project History

• Design / Study Round 3.
• New version based on results of Round 2
• Highly positive user response
• Identified new user population/collection
• Students and scholars of art history
• Fine arts images
• Study Round 4
• Compare the metadata system to a strong,
  representative baseline

43
New Usability Study

• Participants Collection
• 32 Art History Students
• 35,000 images from SF Fine Arts Museum
• Study Design
• Within-subjects
• Each participant sees both interfaces
• Balanced in terms of order and tasks
• Participants assess each interface after use
• Afterwards they compare them directly
• Data recorded in behavior logs, server logs,
  paper-surveys one or two experienced testers at
  each trial.
• Used 9 point Likert scales.
• Session took about 1.5 hours pay was 15/hour

44
The Baseline System

• Floogle
• Take the best of the existing keyword-based image
  search systems

45
Comparison of Common Image Search Systems
46
sword
47
(No Transcript)
48
(No Transcript)
49
(No Transcript)
50
Evaluation Quandary

• How to assess the success of browsing?
• Timing is usually not a good indicator
• People often spend longer when browsing is going
  well.
• Not the case for directed search
• Can look for comprehensiveness and correctness
  (precision and recall)
• But subjective measures seem to be most
  important here.

51
Hypotheses

• We attempted to design tasks to test the
  following hypotheses
• Participants will experience greater search
  satisfaction, feel greater confidence in the
  results, produce higher recall, and encounter
  fewer dead ends using FC over Baseline
• FC will perceived to be more useful and flexible
  than Baseline
• Participants will feel more familiar with the
  contents of the collection after using FC
• Participants will use FC to create multi-faceted
  queries

52
Four Types of Tasks

• Unstructured (3) Search for images of interest
• Structured Task (11-14) Gather materials for an
  art history essay on a given topic, e.g.
• Find all woodcuts created in the US
• Choose the decade with the most
• Select one of the artists in this periods and
  show all of their woodcuts
• Choose a subject depicted in these works and find
  another artist who treated the same subject in a
  different way.
• Structured Task (10) compare related images
• Find images by artists from 2 different countries
  that depict conflict between groups.
• Unstructured (5) search for images of interest

53
Other Points

• Participants were NOT walked through the
  interfaces.
• The wording of Task 2 reflected the metadata not
  the case for Task 3
• Within tasks, queries were not different in
  difficulty (tslt1.7, p gt0.05 according to
  post-task questions)
• Flamenco is and order of magnitude slower than
  Floogle on average.
• In task 2 users were allowed 3 more minutes in FC
  than in Baseline.
• Time spent in tasks 2 and 3 were significantly
  longer in FC (about 2 min more).

54
Results

• Participants felt significantly more confident
  they had found all relevant images using FC (Task
  2 t(62)2.18, plt.05 Task 3 t(62)2.03, plt.05)
• Participants felt significantly more satisfied
  with the results
• (Task 2 t(62)3.78, plt.001 Task 3 t(62)2.03,
  plt.05)
• Recall scores
• Task2a In Baseline 57 of participants found all
  relevant results, in FC 81 found all.
• Task 2b In Baseline 21 found all relevant, in
  FC 77 found all.

55
Post-Interface Assessments
All significant at plt.05 except simple and
overwhelming
56
Perceived Uses of Interfaces
Baseline
FC
57
Post-Test Comparison
FC
Baseline
Which Interface Preferable For
Find images of roses Find all works from a given
period Find pictures by 2 artists in same media
Overall Assessment
More useful for your tasks Easiest to use Most
flexible More likely to result in dead
ends Helped you learn more Overall preference
58
Facet Usage

• Facets driven largely by task content
• Multiple facets 45 of time in structured tasks
• For unstructured tasks,
• Artists (17)
• Date (15)
• Location (15)
• Others ranged from 5-12
• Multiple facets 19 of time
• From end game, expansion from
• Artists (39)
• Media (29)
• Shapes (19)

59
Qualitative Observations

• Baseline
• Simplicity, similarity to Google a plus
• Also noted the usefulness of the category links
• FC
• Starting page well-organized, gave ideas for
  what to search for
• Query previews were commented on explicitly by 9
  participants
• Commented on matrix prompting where to go next
• 3 were confused about what the matrix shows
• Generally liked the grouping and organizing
• End game links seemed useful 9 explicitly
  remarked positively on the guidance provided
  there.
• Often get requests to use the system in future

60
Study Results Summary

• Overwhelmingly positive results for the faceted
  metadata interface.
• Somewhat heavy use of multiple facets.
• Strong preference over the current state of the
  art.
• This result not seen in similarity-based image
  search interfaces.
• Hypotheses are supported.

61
Other Domains

• Applying this to
• Text
• Tobacco Documents Archives
• Medline biomedical texts
• Products/Catalogs
• Dont have a collection would like one

62
Implementation

• All open source code
• Mysql database
• Python web server (Webkit)
• Python code
• Lucene search engine (java)

63
Summary and Conclusions
64
Summary

• We have addressed several interface problems
• How to seamlessly integrate metadata previews
  with search
• Show search results in metadata context
• Disambiguate search terms
• How to show hierarchical metadata from several
  facets
• The matrix view
• Show one level of depth in the matrix view
• How to handle large metadata categories
• Use intermediate pages
• How to support expanding as well as refining

65
Summary

• Usability studies done on 3 collections
• Recipes 13,000 items
• Architecture Images 40,000 items
• Fine Arts Images 35,000 items
• Conclusions
• Users like and are successful with the dynamic
  faceted hierarchical metadata, especially for
  browsing tasks
• Very positive results, in contrast with studies
  on earlier iterations
• Note it seems you have to care about the
  contents of the collection to like the interface

66
Summary

• Validating an approach to web site search
• Use hierarchical faceted metadata dynamically,
  integrated with search
• Many difficult design decisions
• Iterating and testing was key
• Bits and pieces were there in industry
• The approach is being picked up too
• One is very similar now endeca.com

67
Advantages of the Approach

• Supports different search types
• Highly constrained known-item searches
• Open-ended, browsing tasks
• Can easily switch from one mode to the other
  midstream
• Can both expand and refine
• Allows different people to add content without
  breaking things
• Can make use of standard technology

68
Some Unanswered Questions

• How to integrate with relevance feedback (more
  like this)?
• Would like to use blobworld-like features
• How to incorporate user preferences and past
  behavior?
• How to combine facets to reflect tasks?

69
The Flamenco Project Team

• Kevin Chen
• Ame Elliott
• Jennifer English
• Kevin Li
• Rashmi Sinha
• Kirsten Swearingen
• Ping Yee
• http//flamenco.berkeley.edu

70
Thank you!
For more information

• flamenco.berkeley.edu