CONTROL: CLEF2003 with Open, Transparent Resources OffLine

1
CONTROL CLEF-2003 with Open, Transparent
Resources Off-Line

• Monica Rogati
• Computer Science Department, Carnegie Mellon

2
Bilingual Retrieval

• Query Translation Approaches
• Blind Feedback
• Query translation
• Blind Feedback
• Retrieval

3
Bilingual Retrieval

• Query Translation Approaches
• Blind Feedback
• Query translation
• SYSTRAN, commercial MT systems
• Some automatically extracted dictionaries
• Blind Feedback
• Retrieval

4
(No Transcript)
5
Bilingual Retrieval
German Query
Italian Query
GIZA

Lemur
Parallel Corpus
Italian Documents
Results
As denoted by parallel in results
6
A few details

• Parallel corpus European Parliament proceedings
• Stemming Rule-based (U. Neuchatel)
• Lemur as retrieval engine
• Giza as IBM Model 1 implementation

7
Key Idea Keep Weights

• Keep TFIDF weights before translation
• Keep translation probabilities
• Keep post-translation weights

8
Multilingual Retrieval
E2DE
English Query
E2ES
2 Step RSV
E2FR
E2E
Bilingual Retrieval
Merging
Results
9
Official Bilingual Runs
10
Official Multilingual Runs
11
Conclusions?

• SYSTRAN-free approaches OK if acceptable parallel
  corpus found
• Pivot language OK but carry the weights!