Galaxy of News: An Approach to Visualizing and Understanding Expansive News Landscapes Earl Rennison In UIST `94, ACM Symposium on User Interface Software and Technology. New York: ACM Press, 1994.

1
Galaxy of News An Approach to Visualizing and
Understanding Expansive News Landscapes Earl
RennisonIn UIST 94, ACM Symposium on User
Interface Softwareand Technology. New York ACM
Press, 1994.

• Paper presentation by Mark Sharp
• 17610554 Information Visualization, Prof.
  Spoerri
• 11/11/2002

2
Paper Summary

• PROBLEM Accessing and understanding news
  information is not well-supported by the
  information infrastructure.
• VISION An intelligent infrastructure that
  automatically builds the correlations and
  relationships between news articles and
  constructs an environment that allows readers to
  dynamically explore and gain understanding.

3
How does it work?

• Articles have features (metadata) extracted by
  parsing algorithms, then they are clustered by
  ARN (a neural network algorithm) and mapped to a
  3D space layout.
• Nodes keyword hierarchy / headlines / full text
• Zoom in with left mouse button, out with right.
  direct manipulation
• Animation (4D) helps user understand what system
  is doing.
  motion an early/pre-attentive visual cue

4
(No Transcript)
5
(No Transcript)
6
(No Transcript)
7
(No Transcript)
8
(No Transcript)
9
(No Transcript)
10
Model components
Temporal and behavior interaction controls
level-of-detail, user orientation cues,
transition to new views.
Spatial construction can be 2-, 3-, or
n-dimensional uses relationships dynamic
(appropriate for news)..
Relationships designer-specified e.g. temporal
ordering
.
News base not raw data objects and annotations
(keywords, slugwords, location, time, subject,
etc.) manually or automatically derived from raw
data.
11
reading
writing
12
(No Transcript)
13
(No Transcript)
14
Which early / pre-attentivevisual processes are
leveraged?

• Position
• Proximity
• Motion
• Brightness
• Size
• Color

15
What is working?

• Principled (algorithmic) feature extraction and
  clustering.
• Direct manipulation.
• True zooming (seamless exploration of categories,
  document labels, and full texts).
• Dynamic updating of content (new articles).

16
What is not working or clear?

• Clustering based on skinny metadata rather than
  full text vectors.
• Keywords are single words, not terms.
• Relationships?

17
What surprised you?

• Naivete about understanding and media studies.

18
Key Insights what I learned

• Detailed look into the architecture of a true
  large text corpus info viz system with many
  desirable features.

19
What is the key contribution?

• True zooming (seamless integration of all levels)
  is feasible in large text corpora.

20
Take-away messages?What can be generalized?

• Computational feasibility forces some
  compromises.
• What is not working
• Human heuristics (relationships?)
• BUT help is on the way (bigger iron)

21
3 questions for groupand class discussion.

• Is volume and lack of organization really our
  biggest problem with modern news information?
• Would you use Galaxy of News? Why or why not?
• What other kinds of text data would you like to
  see this approach applied to? How might a
  different domain affect the specification of
  metadata object representations and/or
  relationships?

22
TileBars Visualization of Term Distribution
Information in Full Text Information
AccessMarti HearstProceedings of the ACM
SIGCHI Conference onHuman Factors in Computing
Systems (CHI), pp. 59-66, Denver, CO, May 1995.

• Paper presentation by Mark Sharp
• 17610554 Information Visualization, Prof.
  Spoerri
• 11/11/2002

23
Paper Summary

• PROBLEM Traditional IR is focused on text
  databases consisting of titles and abstracts
  assumptions are not necessarily appropriate for
  full text.
• VISION Utilize term distribution within the text
  as well as overall frequency to model document
  relevance. Replace opaque ranking with a
  transparent means for swift appraisal of the
  query-document relationship.

24
How does it work?

• TextTiling algorithm partitions full text into
  adjacent, non-overlapping, multi-paragraph
  segments reflecting subtopic structure based on
  term co-occurrence and repetition.
• Segments are scored for similarity to query
  terms.
• Display shows document length, term frequency,
  and term distribution across segments.

25
Length of rectangle length of
document Each gray square 1 tile
(segment) Tile darkness term
freq. Query term sets tile rows
26
(No Transcript)
27
(No Transcript)
28
Which early / pre-attentivevisual processes are
leveraged?

• Length
• Position
• Darkness (gray scale)

29
What is working?

• Elegant rep. of document length.
• Adjacency of tiles between term rows gt overlap.
• Gray scale leverages relative (vs. absolute)
  judgment.
• Meaningful labels (start of text).
• Direct click link from tiles to text segments.
• Starting TREC/TIPSTER evaluation.

30
What is not working or clear?

• Depends on skillful Boolean query formulation
  (e.g. no stopwords).
• Doesnt appear to be scalable to large queries
  (gt3 conjunctive terms).

31
What surprised you?

• Because they do have a natural visual hierarchy,
  varying shades of gray show varying quantities
  better than color.

32
Key Insights what I learned

• Relevance ranking is not the only game in town
  for putting cognitive cues on multi-document
  retrievals.

33
What is the key contribution?

• Text segmentation can enhance traditional
  (whole-document) IR as well as fact retrieval.
• Novel paradigms for text retrieval can be both
  principled and computationally efficient.

34
Take-away messages?What can be generalized?

• Marti Hearst is a major player in text mining /
  text visualization.

35
3 questions for groupand class discussion.

• Instead of integer term frequency, what else
  could be used to color the tiles for relevance?
• How might documents be ranked?