[[page created automatically from word-processed document; for original see: Postscript version]]

CONTEXT-AWARE

INFORMATION RETRIEVAL

Peter Brown and Gareth Jones

Dept of Computer Science

Univ. of Exeter, UK

Context-awareness as a discipline

• Compare two disciplines: word-processing and context-awareness.
  • Word processing:
    • relates directly to a user need.
    • a loosely defined term.
    • its meaning changes over time.
    • the basics are well understood.
  • Context-wareness:
    • must be a (seamless?) part of a larger application.
    • a precise definition is at the end of the rainbow.
    • the basics are poorly understood.

Focus of the talk

• Two aspects of context-awareness:

  (1)

  Proactiveness.

  (2)

  How context changes.

• We have concentrated on CAR (Context-Aware Retrieval) of information.
• Two established retrieval disciplines:
  • IR (Information Retrieval): one user, different query every time, large number of (fairly static) documents.
  • IF (Information Filtering): many simultaneous users, same query every time, one document at a time.
• Our paper on relationship between IR/IF and CAR is at
  
  Link to a our paper
• Three basic elements:
  • (1) User's current context; perhaps many users.
  • (2) A collection of documents to be retrieved from.
  • [[(3) The recipient, here assumed to be the user.]]

Proactiveness

• Proactiveness can be in element (1) and/or (2). e.g.:

  (1)

  setting all or part of the user's context via sensors.

  (2)

  attaching to a document a triggering condition, which triggers the document when the triggering condition matches the user's context.

• We believe that in most applications proactiveness will need to be done intelligently:

  (1)

  Sensors are too low level.

  (2)

  Algorithms for matching trigger conditions need to be intelligent.

• Basic needs for all types of retrieval:
  • Efficiency.
  • Effectiveness: precision and (more important) relevance.
• Some CAR needs can be met by IR/IF techniques.

The way context changes

• CAR is harder than IR/IF because:
  • Many users.
  • Many documents.
  • Everything is potentially dynamic.
  • Need for continuous updating.
  • Need for speed.
• CAR is harder than the IR/IF "grand challenge"?
• The saving grace is that the current context may be changing gradually and partly predictably. This can be exploited for:
  • Efficiency.
  • Use of caches.
  • Extrapolation (c.f. guessing future cellphone cells).

Conclusions

• For CAR:
  • Proactive behaviour needs sophisticated algorithms, in the same way that IR/IF does.
  • Some applications will only be feasible if change is gradual and partly predicatable.
• Do these conclusions apply in general to context-aware applications?