Kohonen Mapping and Text Semantics

1
Kohonen Mappingand Text Semantics

• Xia Lin
• College of Information Science and Technology
• Drexel University

2
Ten years ago

• Lin, X., Soergel, D., Marchionini, G. A
  self-organizing semantic map for information
  retrieval. SIGIR89, pp. 262-269.
• Applied Kohonens mapping to a small document
  set
• 140 documents
• 25 indexing words

3
A Semantic Map for 140 AI Documents
1
3
1
1
2
1
6
2
4
1
2
search
network
library
online
Citation database
application
2
1
1
2
2
29
2
3
others
Machine learning
knowledge
2
1
1
1
1
6
natural
retrieval
1
1
4
4
1
1
systems
2
1
1
1
process
2
1
2
2
3
language
1
1
1
3
expert
1
6
1
Intelligent
research
3
2
3
3
4
2
1
4
Features of the Semantic Map

• Reveal frequencies and distribution of underlying
  data.
• Preserve metric relationships as faithfully as
  possible while mapping from high-dimensional
  data to a two-dimensional display
• Display co-occurrence structures through its
  neighborhood structures.

5
Why do you prefer using Self-organizing Map (SOM)
to textual information?

• The power of abstraction
• The feature of self-organization
• The format of output -- rich information for
  display

6
Information Abstraction

• SOM utilizes statistical information of text in a
  unique way
• Both individual data and their inter-relationships
  are represented.
• Learning takes place gradually
• To tolerate uncertainty/fuzziness in the input
  data
• It represents large amount of data economically
• Similar to the way the brain processes/stores
  information?

7
Information Organization

• SOM uses the input data to make a random network
  become an organized network.
• Each piece of information will find its own
  identity (the best place) on the map.
• All the related information should be organized
  together.
• A compromise or enforcement of both individual
  responsibilities and social responsibilities.

8
Information Visualization

• SOMs output is an associative network that can
  be used to implement various interactive
  functions of the interface
• A good overview of underlying data
• A variety of topologic structures
• Sizes of groups, distances, weights of vectors,
  patterns of inputs, etc.
• A space of both documents and terms
• Effective use all the space of the
  two-dimensional area.

9
How much semantics are represented in Kohonens
map?

• Its an open question.
• Understanding can be gained through comparisons
  and applications.

10
(b) Principal component analysis
dove
hen
duck
goose
owl
hawk
eagle
fox
cat
dog
wolf
tiger
lion
cow
horse
zerba
(a) Hierarchical cluster
(c) Kohonen's feature map
11
graph
M1
M3
Human
C1
M4
computer
tree
Tree
M2
minor
survey
C4
Interface
M2
time
C5
C2
response
System
C3
user
M1
EPS
M3
computer
Graph
C2
C1
Response time
system
interface human
C3
minors
User
EPS
C4
M4
C5
survey
(a) Display of the Latent semantic indexing result
(d) Document and term map by the feature map
12
Visual SiteMap
13
(No Transcript)
14
(No Transcript)