IR to UIR Future Directions in Information Retrieval

1
IR to UIRFuture Directions in Information
Retrieval

• Rohini K. Srihari
• Department of Computer Science and Engineering
• State University of New York at Buffalo

2
Future of Information Retrieval

• Objective More productive IR sessions by
  combining search with discovery
• Key Challenges
• Response customized to user needs (HARD track at
  TREC)
• Granularity of response
• User preferences
• Genre of response (general news,
  subjective/opinion, background etc.)
• Interactive IR
• Search information extraction text
  mining/discovery follow-up search
• Hybrid Representations for IR
• Incorporate both traditional word models (e.g.
  vector space) and concept/association models
  enabled by information extraction
• Permit text mining

3
CCG for FAA Website
Lightning
AIRPLANE
Fuel Tank
HAZARD
Wiring
AVIATION
Statistics
Fuel Tank Ignition events
Windshear
ACCIDENT
Pumps
Air_traffic_ _control_tower
Ice/snow
Small Bomb
In-flight fire
Previously viewed page(s)
Fatalities
hull losses
Runway Incursions
requested page
UIR module determines that these two documents
reveal new association between wiring and
accidents.
4
Unapparent Information Revelation (UIR)

• Document collections may reveal interesting
  information other than what is explicitly stated
• E.g. interesting links between gene names, genes
  and diseases
• E.g. Links between people and organizations
• Hidden information a consequence of multiple
  sources and authors working independently
• Users, e.g. researchers, analysts need to
  discover links
• Reverse IR system is given relevant documents,
  trying to generate complex query/narrative
• User has already generated a set of interesting
  documents through some searches what complex
  query do these represent?
• Required set of automated tools that will
• expose interesting links, generate corresponding
  document set
• Permit refinement of queries based on above

5
Integrated IR UIR System
6
Components of a UIR System

• Domain map representing major concepts and
  associations
• E.g., UMLS, intelligence knowledge bases
• Automatic extraction of key concepts and
  associations from a target document set
• Concepts include named entities, phrases that are
  subjects/objects of events
• Associations include both named relationships
  such as affiliation, age, etc. Also includes
  un-named relationships based on proximity of
  concepts in a sentence, paragraph, document etc.
• Concept Chain Graph
• Probabilistic network with nodes representing
  concepts, edges representing associations
• Document is viewed as a sub-graph footprint is a
  measure of information revealed
• Intuitive visualization, navigation
• Text mining module operating on the CCG