International Joint Conference on Neural Networks IJCNN 2005

1
International Joint Conference on Neural Networks
(IJCNN 2005)

• A Self-Organizing Map
• for Concept Classification
• in Information Retrieval

Guy Desjardins intellagent_at_vif.com Robert
Godin godin.robert_at_uqam.ca Robert
Proulx proulx.robert_at_uqam.ca
2
Agenda

• Information Retrieval
• IR SOM-NN
• SOM Neural Network Model
• Experiment and Results
• Conclusion - Future

3
Information Retrieval

• Text representation
• Indexing documents ? corpus terms

4
Information Retrieval

• Query Document matching

5
Information Retrieval

• Similarity(Qi, Dj) cosine (qi, dj)

6
IR SOM-NN
7
IR SOM-NN
Documents

• Concepts

8
IR SOM-NN

• Documents ? Concepts
• document weights
• tf translation
• idf translation
• concept weights

9
SOM Neural Network Model

• Input layer document vectors (t1, t2,..., tn)
• Output layer document clusters c1, c2, ..., cj
• Need to manually translate into term clusters

Standard Topology in Information Retrieval
Document clusters
Documents
10
SOM Neural Network Model

• Term-document matrix

11
SOM Neural Network Model
Our Model

• Input layer term vectors (d1, d2, ..., dm)
• Output layer Concepts C1, C2, ..., Cj

Concepts
Terms
12
SOM Neural Network Model

• Learning rule

13
Experiment

• Collection TREC FT943
• 17 109 documents
• 15 queries
• 71 011 index terms
• SOM parameters
• two-dimensional map 20 ? 20  400 nodes
• learning rate ? 0.1
• sacling factor ? 3750
• maximum starting neighbourhood ? 3

• Averages
• 7 documents / index term
• 29 terms / document

14
Results

• 362 output nodes activated (90)
• Average of 196 terms / node
• Center of map high density nodes
• nodes 9-11 9-11 clustered 36 of all terms
• Periphery of map more specific concepts
• number of terms / node ? 30

15
Results
16
Results

• Information retrieval

17
Conclusion - Future

• SOM Meaningfull concept clusters
• IR Full term vectors ?? Concept vectors
• Is SOM adequate for IR ?
• Dimensionnal reduction (71 000 to 400 200X)
• Future work
• Recursive SOM
• Explore sensitivity of the parameters
• Compare with other concept approaches