Keizai Crosslanguage Retrieval, Summarization

1
Keizai - Cross-language Retrieval, Summarization
Translation

• Jim Cowie, Mark Davis, Eugene Ludovik, Sergei
  Nirenburg, Bill Ogden, Hugo Molina-Salgado, Nigel
  Sharples
• Computing Research LaboratoryNew Mexico State
  University

2
Talk OverviewGetting information from text you
cannot read

• Human Factors issues for CLTR
• Two studies
• Visualization for document selection
• Interactive CLTR
• Keizai design

3
CLTR users

• The advanced foreign language reader
• needs help with query writing
• The monolingual
• What are their goals?
• question answering,
• finding relevant document for full translation
• finding a relevant link
• ??

4
Two human factors

• Help with judging the relevance of documents you
  cannot read
• Help with query formulation

5
Document Thumbnail Views

• Single glance keyword information
• instances collocations
• frequencies distributions
• missing keywords
• Familiar document format enables
• easy interpretation
• locating relevant text
• Answers why was this document returned

6
The J24 Interface
7
TREC Aspectual Recall Task
Seconds
Read topic
8
Document Selection Task
Seconds
Read topic
9
Method

• 4 college student participants
• Used both thumbnails alone and titles alone.
• Judged 20 documents for relevance to each of 11
  selected TREC-7 topics
• All query terms did not appear in all documents.

10
Document Selection Study
11
(No Transcript)
12
Conclusions for document selection study

• Keyword occurrence information can be more useful
  than document titles
• In cases where keyword distribution can determine
  relevance
• Expand beyond query terms and keyword

13
Other methods

• Tile Bars

• Winquery

• Web Conventions

14
(No Transcript)
15
Other methods

• Tile Bars

• Winquery

• Web Conventions

• Aspect summaries

demo
16
Arctos Interactive Cross-Language IR

• English query
• Exploit WWW resources to help improve query
• Use URSA and J24 to retrieve documents
• WWW MT

Arctos demo
17
Arctos interactive study

• One non-German user formed English queries from
  22 TREC-7 CLTR topics.
• Iterated query and judged the top ten German
  documents (8 hours).
• Compared judgements with NIST

18
Results
User Judgements
not relevant
relevant
relevant
69
1
NIST Judgements
not relevant
11
43
19
Keizai

• Integrates various components in a prototype tool
  for analysts.
• URSA Unicode Retrieval System
• MINDS Multilingual Documents Summarization
• Correlli Rapid MT development system.

20
MINDS

• User configurable summarization system based on
  sentence selection
• Can focus on query and/or named entities
• Summarizes documents in Spanish, Japanese,
  Russian, Turkish, Korean, and English

21
Human aided query translation

• English terms are back-translated to generate
  foreign candidates
• Pre-search removes non-occurring query terms.
• Natural disambiguation and query expansion.

22
demo
23
(No Transcript)
24
Related work

• Mulinex screens
• Query assistant
• Translated summaries
• MuST
• CLTR w/English summaries
• And other...

25
(No Transcript)
26
(No Transcript)
27
Planned Research

• Improved integration of Japanese-English
  Korean-English MT
• Development and evaluation of a variety of user
  interaction modes.
• Are user-aided queries as good?
• Can users judge relevance from English summaries?

28
Interactive CLTR

• User-aided query translation
• User can help disambiguate query terms
• Query expansion occurs naturally
• Judging relevance - when text is unreadable!
• Show English keyword summary aspects in result
  list
• Summarize key passages in English