CrossLingual Information Retrieval EnglishTamil

1
Cross-Lingual Information Retrieval
(English-Tamil)

• Proposers Dr. T.V.Geetha
• Dr. Ranjani Parthasarathi
• Department of Computer Science Engg.
• CEG, Anna University
• Chennai 600025

2
Proposed Components Resources

• Components
• Morphological Analyser
• Keyword translator
• IE engine (POS,Shallow Parser, NER, UNL)
• English Tamil Transfer (UNL/Template)
• Summary Generator

• Lexical Resources
• Domain-specific Enhancement of Bilingual
  Dictionary (English-Tamil)
• Domain-specific Parallel corpus

3
Components
4
Morphological Analyser (Tamil)

• Task Enhancement of existing MA
• Technique Used Rule-based
• Expected performance
• gt90 for the chosen domains (standard corpus)
• Domains Chosen
• Popular science
• Evaluation metrics
• Cross-check robustness with standard corpus/MA

5
Keyword translator

• Task Translation of keywords with sense
• Technique Used
• Statistical based and Language model based word
  sense disambiguation
• Expected performance
• gt 90 for the chosen domains
• Domains Chosen
• Popular science
• Evaluation metrics
• Accuracy and precision (Corpus-based evaluation)

6
Information Extraction for Tamil

• Task Extract info from search results into
  Template / UNL representation
• Involves POS, Shallow Parser, Named Entity
  Recognizer, Template matching, Tamil to UNL
  enconverter
• Techniques Used
• Statistical and rule based
• Expected performance
• gt 85 precision and recall for the chosen domains

7
English Tamil Transfer (UNL/Template)

• Task Transfer info from English search engine
  results in Template / UNL representation to Tamil
  
• Involves Parallel corpus based matching / UNL to
  Tamil deconverter
• Techniques Used
• Statistical and rule based
• Expected performance
• gt 85 for the chosen domains

8
Summary Generator

• Task From Template to Summary
• Involves Morphological generation, multi-sentence
  fusion
• Techniques Used
• Statistical and rule based sentence generation,
  and reference resolution
• Expected performance
• gt 85 precision for the chosen domains

9
Lexical Resources
10
Domain-specific Enhancement of Bilingual
Dictionary (English-Tamil

• Existing Dictionary (50000 root words generic)
  to be enhanced
• 20000 domain specific words to be added from a
  chosen corpus of 15K sentences
• Expected size 70000 root words
• Other Indian languages 20K-50K words (being
  developed)
• Evaluation metrics Corpus-based evaluation

11
Domain-specific Parallel corpus

• Expected outcome 15K sentences in the chosen
  domain fully tagged (manually/ semi-automatically)
  in both languages
• Not available for many languages
• 10K sentences at the end of phase 1, and the
  remaining 5K at the end of phase 2
• Evaluation metric Accuracy
• Indirect metric precision of search results

12
Thank you