Jean-Charles LAMIREL, Jieh HSIANG

1
Using a Background Neural Model in a Digital
Library

• Jean-Charles LAMIREL, Jieh HSIANG
• Liu WJ

LORIA, Nancy, France
2
The CORTEX team

• Research areas Biological-like models for
  intelligent information management
• Applications
• Autonomous robotics and in-board intelligence
• Numerical classification (vs. symbolical)
• Information retrieval and discovery

3
The CORTEX information retrieval and discovery
activity

• Main themes of research
• Interface for personalized access to information
• Intelligent multimedia data mining
• Web - Documentary database interaction
• Collaborations
• ORPAILLEUR INRIA team, INIST, LaVillette, NSC
  Taiwan, industry...
• European projects SCHOLNET, EISCTES

4
Some examples of application

• Adaptive environment for assistance to
  investigation on the Web
• Multi-topographic navigation MultiSOM
• For multimedia data mining
• For data mining on full text (patents)
• Numerical-symbolic collaboration

5
Presentation summary

• Introduction
• Basic set of functionalities for information
  discovery
• Limitations of the classical methods for
  information discovery
• The MultiSOM model Butterfly application
• Basic behaviour
• Extensions
• Management of textual information

lamirel_at_loria.fr
6
Basic set of functionalities for information
discovery

• Synthetical view of the studied domain
• Distribution of the thematical indicators of the
  domain
• Highligting of regularities / weak signals
• Management of several type of synthesis
• Interactivity
• Dynamic data mixture / type of need
• Choice of meta-orientation of investigation
• Setting of the granularity level of the analysis
• Multimedia

7
Managing different kinds of queries for discovery

• Exploratory (no goal)  Which is the contents
  of the database ?
• Thematic (general orientation)  Images of
  space conquest 
• Connotative (hidden goal, indirect research)
   Impressive images on human technology 
• Precise  Images of Amstrong moonwalk, July 69 

8
Limitations of the classicalmethods for
information discovery

• Overall view of the studied domain
• Noise
• Complex interpretation (hidden information)
• Local views necessarily independant
• Weaks signal difficult to highlight
• No interactivity
• Passive classification
• Predefined ways to access to information

9
Neural methods for information cartography

• Topographic learning (SOM)
• classification
• projection
• Multi-viewpoint modelization capabilties
  (MultiSOM)
• Intuitive auto-organization of information
• Active maps (IR Navigation)
• Low human intervention during construction
• Multimedia capabilities

10
Butterfly museum application

• Different kinds of query
• Query by keywords
• Query by example
• Different kinds of criteria
• Colour (automatic)
• Shape (manual)
• Texture (manual)
• Problems
• Hand-made classifications
• Combination of results coming from different
  criteria

Yellow very strong,Red not,Edge
strongSpot middle,
11
Butterfly application automation
Global and/or cross viewpoints classifications
User interface
Combination of results
User interface
Validation of insertion or classification
recalculation
Butterfly application
Viewpoint classifications
12
Basic topographic map building

• Data description
• Document (image) index vector eg vector of
  characteristics
• Weighting of the characteristics modalities (very
  strong1, )
• Optionnal IDF weighting (weak signals detection)

13
Basic topographic map building

• Map predefined parameters settings
• Number of neurons
• Structure eg 2D grid with square neighbourhood
• Competitive learning

14
Selection of the winning neuron
Influence on the neigbourhood
Competitive learning
15
Map labelization and zoning

• Map labelization
• Based on the best components of the profiles
• Class or member-oriented
• One single method is not sufficient
• gt Gives an overview of the detected themes
• Map zoning
• Based on the SOM topographic properties
• Based on the best components of the class
  profiles
• gt Gives an overview of the weights of the
  themes

16
(No Transcript)
17
The MultiSOM model
18
Map on-line generalization

• Goal
• Synthethize the map contents by decreasing the
  number of neurons (classes)
• Constraints
• Preserve the map topographic properties
• No classification re-computation
• Method
• Exploitation of the neighbourhood relations on
  the map

19
Map on-line generalization
20
(No Transcript)
21
Semantic viewpoints

• Subspace of the description space
• Can be a field, a subset of keywords, ...
• Possible overlapping sets
• Concurrent or complementary viewpoints
• gtExamples indexer keywords, title keywords,
  authors, , visual characteristics, sounds
• gtButterflies color, shape, texture,

22
Inter-map communication

• Goal
• Cope with the limitations of a global map
• Allow communication between viewpoints
• Constraints
• Interpretable behaviour
• Method
• Re-projected data Transmitters neurons
• Two steps
• 1) Activation of a source map (directly or
  through a query)
• 2) Transmission to target maps

23
Inter-map communication
24
Inter-map communication

• A function
• Two modes
• Possibilistic (weak thematic relations over
  viewpoints) 
• Probabilistic (mesure of the themes similarities)

gt g class belonging degree
25
Activity coherency
STRONG FOCALIZATION
WEAK FOCALIZATION
26
Inter-map communication
BUTTERLIES
27
Compliance with IR operations
Response NO
Response YES
Question Are there butterflies with spots AND
veins ?
28
Remaining problems (to be solved)

• Validation of the automatic classification
  results by the experts
• Testing of different results merging methods
• Test the use of prototype features in
  classification
• Realization of a Web interface for the maps
• Compare map build-in result combination mechanism
  with external combination mechanism
• Test map capabilities for the help in adding new
  individuals
• Introduce textual data and combine it with images

29
(No Transcript)
30
Experimentation on patents (texts)

• Goal Intelligent technological survey
• Full text analysis of the patents
• Domain of oil engineering
• Provide answers to questions like
• 1. Which are the relationships between patentees
  ?,
• 2. On which specific technology does a patentee
  work ? Which are the advantages of this specific
  technology ? For which use ?,

31
Basic experimental protocol
DILIBReformating
PatentsDatabase
Patents in XMLFormatStructured by Viewpoints
Nominal groupsExtraction
ValidatedMulti-indexes
Interactive maps for analysis
MicroNOMADMultiSOM
lamirel_at_loria.fr
32
Nominal groups extraction

• 1) Lexicographic analysis (compound terms)
• 2) Normalization
• Ex  oil fabrication  and  oil engineering
  gt  oil engineering 
• Results

33
Patents reindexing

• Selected Viewpoints title, use, advantages and
  patentees

34
Example of dynamic analysis
DYNAMIC DEDUCTION Parentee TONEN CORP.  is a
specialist of lubrification of the  automatic
transmission . It products mainly oils based on
 organo- molybdenum compound  whic have the
specific property of having a  friction
coefficient stable stable on a wide range of
temperature 
35
Classical methods (AK-means)
CLASSES MAP
36
Conclusion

• Different viewpoints yield complementary results
• Ex Indexer keywords Closed themes, Title
  keywords Open themes, ...
• Detection of indexation inconsistencies
• Projection of thematic pertinence of a query
• Bilateral synergy images ltgt textual information
• Very rich and flexible inter-map communication
  mechanism
• Cross analysis between viewpoints, dynamics
• No limitation regarding viewpoints type and
  number

37
Perspectives

• Sophisticated 2D mapping, 3D mapping
• Pure image mosaic navigation
• Automatization of communication between
  viewpoints
• Interaction with Gallois lattice map zoning and
  generalization, rule mapping, lattice entry
  points selection
• Applications

• La Vilette interactive browsing through museum
  collection, setting up of exibitions
• INIST Cartography of the Web (EISCTES EEC
  Project)

38
3) Combining Symbolic and Numeric Techniques for
DL Contents Classification and Analysis

• Jean-Charles LAMIREL,
• Yannick TOUSSAINT (Orpailleur)

39
Introduction

• Combining numerical and symbolic methods
• MicroNOMAD Self Organizing Maps (SOM)
• Basic SOM topographic properties
• MicroNOMAD multi-map communication process
• Lattice
• Formal properties and symbolic deduction
• Hierarchical structure and inheritance of
  properties
• Study of projection of SOM over lattice
• Making explicit formal properties on the map
• Map intelligent zoning and labelization

40
Galois lattice

• Symbolic hierarchical method (i1, i2, p1, p2,
  p3)
• Partial order defined by the subsumption relation
  over the set of formal concepts
• (I1, P1) ? (I2, P2) ? I1 ? I2,
• (I1, P1) ? (I2, P2) ? P1 ? P2,
• ? I1, I2 there is a unique meet and join.
• Inheritance of properties
• Extraction of association rules
• Search Engine ? Web, IR

41
I i1, i2, i3, i4, P AI, Robots, Search
Engine, Web, IR i1 Web, IR i2 Web, IR i3
Web, IR, Search Engine i4 AI, Robots
i1, i2, i3, i4 , ?
i1, i2, Web, IR
i4, AI, Robots
i1, i2, i3, Search Engine, Web, IR
?, IA, Robots, Search Engine, Web, IR
R1 Search Engine ? Web, IR
42
Complementarity of approaches

• Kohonen SOM
• Complex weighting scheme
• Difficulty for precise interpretation
• Good illustrative power (topographic structure)
• Good synthesis capabilities
• Non linearity
• Lattice
• High number of classes
• Memory and time consuming
• Hierarchical structure
• Rule extraction
• Incrementality

43
3-steps methodology
Projection
Grouping
Agglomeration
44
Conclusion

• Cosine method seems to be the best of the test
• Good accuracy
• Well-balanced agglomeration
• Agglomeration preserves closed areas on SOM
• Other projection and agglomeration methods have
  to be tested
• Preservation of partial order and inheritance

45
Perspectives

• Evaluation on large corpus Expert
• Rule management
• class quality evaluation
• class labelisation
• Deduction validation on communicating maps
  (lattice extensions)
• Implementation of an operational prototype

46
Other approaches

• Multi-classificator cooperation (PhD)
• SVM
• Stigmergy
• Genetic
• Neural maps
• On-line learning of user s behaviour,
  intelligent relevance feedback

47
Annexes

• Topographic inconsistencies
• Area computation
• Inter-map communication
• Activity coherency

48
Topographic inconsistencies
NO INCONSISTENCIES
WEAK INCONSISTENCIES
STRONG INCONSISTENCIES
49
Topographic inconsistencies
GLOBAL
STRONG
Neuron neighbourhood
50
Area computation
WHILE
SO AS
IN
DO
END DO
51
Inter-map communication
52
Viewpoint oriented Patents Analysis

• Selected Viewpoints title, use, advantages and
  patentees

lamirel_at_loria.fr
53
Themes extending oil live  and  black sludge
control are strongly linked together because
they are neighbours on the map
 black sludge  apparition has a negative
incidence on the  friction coefficient of oil
MAP OF VIEWPOINT ADVANTAGES