Telugu-English%20Dictionary%20Based%20Cross%20Language%20Query%20Focused%20Multi-Document%20Summarization

1
Telugu-English Dictionary Based Cross Language
Query Focused Multi-Document Summarization

• IIIA-2006

Prasad Pingali, J. Jagadeesh, Vasudeva Varma
International Institute of Information
Technology, Hyderabad, India pvvpr,
vv_at_iiit.ac.in July 08, 2006
2
Cross-Language Information Systems

• CLIR, Factoid based Question Answering were
  thoroughly researched
• TREC, CLEF and NTCIR are some of the focused
  workshops in this area
• Efforts are on to make the state-of-the-art
  research usable in applications
• Machine Translation capabilities are a bottleneck
  for end-user application
• CL query focused summarization not studied.

3
CL Query Focused Summarization Use Case
4
Potential Benefits of CLQ Summarization

• A bridge between CLIR and MT
• Output is a paragraph, coherent, readable and
  syntactically correct.
• MT systems need syntactically correct sentences
  for proper translation
• Summary can be tailor made to suit the
  constraints of MT system
• For example, summary with minimum NL ambiguity
  etc.

5
Problem Statement

• to synthesize from a set of 25-50 documents in a
  language L2 that are related to a given topic, a
  brief, well-organized, fluent answer to a need
  for information given in a language L1 , that
  cannot be met by just stating a name, date,
  quantity, etc.

6
Extraction Based Summary

• Extraction based summarization process
• Identify Sentences from documents
• Score these sentences using some function
• Choose the top scoring sentences
• Identify and eliminate exact duplicates and near
  duplicates (redundancy)
• Concatenate the sentences in some logical order
  (e.g., timestamp etc.)
• Sentence scoring function
• Query independent or
• Query based or
• Combination

7
Sentence Scoring using RBLM

• Relevance based language modeling (RBLM)
  framework

• Using conditional sampling, joint probability
  can be re-written as

• Term dependencies in a CL setting calculated as

8
Sentence Scoring using RBLM (contd.)

• P(qi/ej) as translation probability
• Explored in statistical machine translation and
  CLIR
• P(ej/w) as post-translation query expansion
• Calculated using dictionary based
• Term co-occurrence statistics
• Pseudo-relevance feedback

9
Calculations in our Experiments (Telugu-English)

• P(qi/ej) as translation probability
• Used bilingual dictionary and assumed uniform
  probability for all possible translations
• P(ej/w) as post-translation query expansion
• Used Hyperspace Analogue to Language (HAL)
  feature to compute term dependencies
• HAL works based on skip-bigrams, window length
  is a parameter

10
Sentence Score

• The score of each word 'w' w.r.t to the query
  can be written as

• The score of each sentence 'S' w.r.t to the
  query can be written as

P
11
Summary Generation

• Top ranking sentences are tested for redundancy
• Cosine similarity between sentences (unigram
  overlap) used for calculating redundancy
• Candidate sentences after removing redundant ones
  are concatenated in the order of the publication
  date of the document
• Sentences belonging to same document are
  concatenated in the order of occurrence in the
  document.

12
Evaluation

• Experiments in Telugu-English language pair
• Used DUC 2005 dataset
• Manually translated DUC queries into Telugu
• Used Telugu queries as input to system
• System generates English summary
• Summary evaluated against DUC model summaries
  using ROUGE package
• ROUGE package contains many metrics, for
  summaries of paragraph length ROUGE-2 and
  ROUGE-SU4 correlate best with human evaluations

13
An Example
14
Evaluation Results
15
Evaluation Results (per topic)
16
Conclusions

• We extended our mono-lingual summarization
  framework to a cross-lingual setting in RBLM
  framework
• We designed a cross-lingual experimental setup
  using DUC 2005 dataset
• Experiments were conducted for Telugu-English
  language pair
• Comparison with mono-lingual baseline shows about
  90 performance in ROUGE-SU4 and about 85 in
  ROUGE-2 f-measures

17
DUC-2006
18
Invitation
Workshop on Cross Lingual Information
Access http//search.iiit.ac.in/CLIA2007 January
06. 2007 held at IJCAI 07 in Hyderabad, India.
19
Thank you

• Questions?