Providing the foundations for efficient and effective context-aware retrieval

Gareth J. F. Jones (Principal Investigator)
and
Peter J. Brown

Department of Computer Science, University of Exeter, Exeter EX4 4QF, UK
g.j.f.jones@exeter.ac.uk
and
p.j.brown@exeter.ac.uk

DRAFT FOR EPSRC FORM

Investigators

Dr. G.J.F. Jones

Title

Foundations for a context-aware retrieval engine. [[Or Foundations for building context-aware retrieval engines.]]

Start and duration

Feb 1st 2001 for 4 years (to allow research studentship to be offset by up to a year)

Objectives

The key aspects of performance in context-aware retrieval are (1) precision: delivering information that is really relevant to the user; (2) speed (especially because context changes rapidly in many applications); (3) ability to run on a hand-held device starved of computational and communications resources.

Our objective is, within the discipline of a best-match IR approach, to improve performance by:

  1. finding good algorithms for scoring and weighting retrieved documents and their sub-fields.
  2. finding good algorithms for converting a context into a query, and assigning weights to its fields; finding algorithms to post-process the retrieved outputs to meet the user's personal needs.
  3. performing field tests to evaluate the results of at least the first two of the above, and as a result to improve our algorithms. (We plan that the field tests will be part simulated and part real.)
  4. covering both proactive (triggered by the data) and interactive (initiated by the user) retrieval, and also covering both physical context and context representing user's personal needs.

Summary

Context-aware Retrieval (CAR) is concerned with retrieving information according to the user's current context. Typically the user is mobile, and their context is derived from a rich variety of sensors (e.g., for location, GPS or phone cells) or from direct settings (e.g. user personalisation). A context consists of a set of fields, where the fields are derived directly or indirectly from sensor values. CAR has some characteristics derived from traditional Information Retrieval and Information Filtering (and also from Databases), plus some unique characteristics of its own, for instance those resulting from the need for continuous/incremental retrieval. CAR will, we believe, achieve prominence when the current commercial position-based services ("Tell me the nearest Macdonalds") evolve into second generation systems, with richer contexts and wider information sources.

At present, apart from a few groups in the USA, CAR is a largely unresearched area. We have published certain ideas that we hope we will provide foundations for building successful CAR engines. These ideas include: context-aware caching; use of a context diary to record history and likely future; the deriving of a `context of interest' from the user's current context. These techniques work best when the user's context is changing gradually and semi-predictably -- as is the case in most real applications. The purpose now is to develop ideas and to prove (or disprove) their viability. We can build on our existing prototype implementation, which has already produced some (limited) results. The outcome could be a basis for the commercial CAR systems of the future.

Beneficiaries

We hope that the final results of the work will be useful to companies offering p-commerce products; indeed our dream is that the work will be the foundation of a new generation of such products, offering the opportunity for wider contexts and wider data sources, and, best of all, delivering information that is really what each individual user wants. Perhaps more importantly, our work is of value to information providers, as it provides a highly focussed means of publicity: this could be vital to small enterprises in rural areas who want to attract visitors.

Notes for Case for Support

see draft case for support.

Staff

Two named RAs: Peter Brown (2/5 time) and Lindsey Ford (1/2 time).

One project studentship (to start 6 month to a year after the project start); maintenance fee 22.3K.

Travel and subsistence to cover IR and mobile computing conferences and field trials.

Consumables: 2-3 PTDs (= mobile phone + PDA), 2-3 GPS, software licences, paying field test subjects, general field test costs, payment to department for research student?.

Exceptional items: 8,220 for session fees for student; phone charges for PTD, including field trials.

2K for employing helpers to create ideal test set results.

Equipment: two laptops, desktop PC for research student. Todo: think about field trials and their equipment needs.

Collaboration

Already agreed: Trilogy, Xerox. Todo. Possibilities: SW tourism, HP (via Mik Lamming link -- BUT will be in USA, not Bristol).

Referees

Possibilities: Robertson @ Microsoft, Lamming @ HP Palo Alto (mik@hp.com), Prof. Nigel Davies @ Lancaster (nigel@comp.lancs.ac.uk, Linington @ UKC (pfl@ukc.ac.uk), Gareth's suggestions.