International Joint Conference on Neural Networks IJCNN 2006

1
International Joint Conference on Neural Networks
(IJCNN 2006)

• An Auto-Associative Neural Network
• for 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 NNs
• Our A-A NN 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 NN BAM Bi-directional Associative Memory
7
IR NN MLP Multi-Layer Perceptron
8
IR NN SOM Self-Organized Map
9
Our A-A NN modelAuto-Associative Neural Network
10
Performance issue

• Number of calculations per document
• Corpus of 26 000 terms 2,7 ? 109
• Corpus of 100 000 terms 4,0 ? 1010
• Dynamic activation of connections
• For active neurons only 4000
• Reduced by a factor of 6,7 ? 105
• Local inactive neurons in a specific document do
  not interact with active ones

11
Hebbian learning rule
Anti-hebbian factor applies one iteration on
two Bégin Proulx 96
12
Experiment

• Test collection
• 2 000 documents
• 7 queries
• Extracted from TREC FT943
• Corpus 25 838 index terms
• Total of 20 relevant documents

13
Results

• Precision-recall curves

14
Results

• Rank of relevant documents

15
Results

• Global metrics

16
Conclusion - Future

• Benefit
• Retrieves the leading relevant documents faster
• Drawback
• Still converges to fast toward the attractor
  vectors
• Future
• Test with larger collection
• Recursive approach

17
Questions ?