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

EXPLOITING CONTEXTUAL CHANGE

IN CONTEXT-AWARE RETRIEVAL

Peter Brown and Gareth Jones

Dept of Computer Science

Univ. of Exeter, UK

Context-aware retrieval (CAR)

• A new application area for retrieval. Some examples:
  • Proactive tourist guides; context detected by sensors; triggering condition (query) supplied by information provider.
  • Interactive tourist guides, where query is enhanced by user's context.
  • Proactively adding links to document on user's screen; context is computational one.
  • Exploitation of personal preferences, history.
  • Office applications geared to role and activity.
  • Reminders, memory aids.
  • Commercial services: `Tell me the nearest Macdonalds'.

CAR concepts

• Needs expertise from retrieval and mobile computing and ... .
• User's current context is typically a collection of fields, covering various data types (text, location, time, image, etc.).
• CAR is related to IF/IR, but not the same.
• Continuous retrieval as context changes.
• Retrieval may be intrusive: high precision is vital; may decide to deliver nothing at all.
• The richer the context, the better the retrieval.

CAR matching

• Retrieval involves a match of user's context, field by field, against context associated with each document.
• Field values may be ranges rather than points.
• Three research areas for improving precision:
  • algorithms for matching individual fields, especially non-text fields.
  • algorithms for weighting fields.
  • algorithms for accumulating overall scores.
• ... BUT evaluation is hard.

Context change

• CAR is hard relative to IF/IR; one compensating advantage may be that context changes gradually and semi-predictably.
• To record context change we build a Context Diary; it covers past and future.
• We use the Context Diary to predict the Context-of-interest, which is just `ahead' of the current context. This can be used to improve speed and precision.
• Context-aware caching:
  • predict the union of a user's current contexts over the next, say, 15 minutes. This union consists of context fields whose values are ranges.
  • retrieve the documents that best match the union, and use these as a cache.
  • cache will improve speed, and cater for disconnected operation.
  • works best when small changes in context lead to small changes in what is retrieved.

Using context change to improve precision

• Possible approaches:
  • a changing field should have more weight than a currently static one.
  • rate of change is significant.
  • relevance feedback can also affect weights.

Current experimental system

• Retrieval engine written in Java; allows considerable tailoring, e.g. supplying a new matching algorithm for Location.
• Use of pipeline architecture, e.g.:
  • calculate user's context ->
  • massage user's context ->
  • proactive retrieval ->
  • interactive retrieval to factor in user preferences ->
  • massaging final scores, using the context diary.

Summary

• High-performance CAR is a need for many future mobile applications.
• CAR needs high precision, fast speed (even with fully dynamic data).
• Analysis of context change provides a means of achieving these needs. Techniques include:
  • recording a context diary.
  • predicting the context-of-interest.
  • using a context-aware cache.