1. Summary of research

As explained in my interim report, my circumstances have changed since I applied for the Fellowship. The two main points ate:

• my former research group at the University of Kent has now spread out, with only one person remaining at Kent.
• I have been working at a Research Fellow at the University of Exeter, and have involved a new (and much younger) colleague, Gareth Jones, in my work; Dr. Jones is an expert in retrieval.

Overall these changes have been positive, and have given a wider base and a wider set of intellectual inputs to my research.

1.1. The research programme

The research has, as planned, centred on computer applications that are context-aware. An example of such an application is one to aid a tourist: the tourist carries a mobile computer with attached sensors to determine their context (for example their location may be determined by a GPS sensor), and information about nearby tourist sites is automatically presented on the computer screen as they move from one place to another.

As sensors become cheaper and more ubiquitous, they can detect an ever richer context, e.g. location, temperature, orientation, objects nearby, computer state, etc. Moreover the sensor-detected context can be augmented or overridden by a virtual context set by the user; for example they might set some preferences, or override their true location by a location they want to pretend to be at.

Context-aware applications already exist in a limited form, for example some mobile phone companies provide an application that tells you the nearest hotels to your current location. However, in order to realise their potential, they need to cater for richer contexts and wider sources of information.

A key conclusion, early in my research programme, was that the methods that context-aware applications use for retrieval of information do not scale up beyond small contexts and limited information sources. New aproaches are needed to meet the needs of context-aware retrieval (CAR), and I decided to focus the research programme on this need; this focus was more than envisaged in the original proposal.

The first step, done in close collaboration with Dr. Jones, was to look at two established fields:

• Information Retrieval (IR), as exemplified by web search engines.
• Information Filtering (IF), as exemplified by certain notification services.

Both these fields have made immense strides over the last thirty years -- it is still a wonder to me that a web search engine will search a billion pages in a negligible time -- and CAR can potentially gain from this. We studied the relationship of CAR to traditional IR and IF, and published two papers [3, 4] on this. In simple terms our conclusions were: (1) CAR is a mixture of IR and IF plus a bit extra; (2) giving good performance in CAR, in terms of speed and relevance of the information found, is harder than in IR or IF; (3) to compensate for (2), we need to find some particular advantages of CAR that can be exploited.

A more specific conclusion was the the IR concept of `best match' retrieval needed to be adopted in CAR. This concept is that, instead of a `yes or no' Bollean retrieval, each potential document for retrieval should be given a score giving how well it matched thew user's needs; this score is derived from weighting various contributory factors -- for example in CAR one factor may be the location, and the weight of this might be changed according to the user's circumstances.

In order to investigate new approaces to CAR, I decided to construct a new retrieval engine, based on the previous `stick-e note' work described in my Leverhulme application, but:

• using best-match retrieval.
• allowing flexibility in weightings and in matching algorithms. The idea is that any researcher can plug in a new weighting sceme or a new matching algorithm, and test its performance. One such matching algorithm might be concerned with matching two locations, and would incorparate ideas on how the value of information about tourist sites decays as the site becomes further from the user.
• covering, in a uniform way, a wide range of contextual elements, both real and virtual.

I did not personally work on this implementation, though I specified it. Instead I spent the bulk of the programme exploring new ideas to improve the performance of CAR. I believe that one advantage of CAR that we can exploit is that the user's context normally changes slowly and semi-predicably. For example the location of a tourist will normally change slowly, and future locations can be predicted -- though sometimes the prediction will be upset by a change in direction. I have designed new matching and weighting algorithms to exploit this (for example a location just ahead has more weight than a location behind. I have also designed caching techniques that try to build a cache of all the documents the user may need during the next 10 (say) minutes. A cache is much smaller and faster to search than a full document collection, and has further advantages on mobile devices that are only periodically connected. Together with Dr. Jones, I have produced a paper [1] on this work, and this has been accepted for an ACM conference in May 2002.

1.2. Subsidiary work

The CAR work has represented about three-quarters of my time (a higher proportion than envisaged), but there has also been subsidiary work. Firstly, I have dome some work, involving Exeter students, on user interfaces for context-aware work. So far this work has not replicated the success of the CAR work.

A second, more successful, activity has been working with staff at Southamton Unicersity (where I have become a visting Professor) on fundamental studies of hypertext, and, more recently, how hypertext retrieval relates to CAR. We have published one paper [2] and have two in preparation.