Research issues in context-aware retrieval: characteristics of context-aware retrieval

Peter Brown

Department of Computer Science, University of Exeter, Exeter EX4 4QF, UK
P.J.Brown@ex.ac.uk

ABSTRACT

We summarise some of the characteristics of context-aware retrieval that differentiate it from traditional Information Retrieval and Information Filtering. This note may be read as an introductory background, before reading the other notes in this series.

A typical example

A typical example of a context-aware retrieval (CAR) application has a user carrying a number of sensors; the user moves around, and, as she does, retrieves information that is relevant to her current context. The current context can be derived directly from the sensors or other available data, or it can be synthesised from this low-level data. For example a "clever" synthesiser may try to use sensor values to work out whether the user is working alone, in a meeting, in communication with a remote person, etc.; it can then set an "activity" field in the current context to reflect this. At a lower level a synthesiser may combine information from many location sensors to create a "near-to" field, that names the objects currently near to a given object. In addition the current context may contain "personal preferences" fields that are set directly or indirectly by the user.

This example gives a flavour, and serves to introduce the more general characteristics that are outlined below.

Characteristics

Some characteristics of context-aware applications are:

  1. retrieval involves the current context and a document collection. Often each document in the collection will have a context associated with it (for example if the document relates to a tourist attraction this context may involve a location, some times representing opening hours, and perhaps a temperature range -- static outdoor activities, such as sitting in a outdoor restaurant will be unattractive if it is cold). To make retrieval easier contexts attached to documents should take the same form and use the same name-space as the current context.
  2. contexts consist of a set of different fields; often, as far as retrieval is concerned, these fields can be treated as independent of each other.
  3. each field has an associated datatype, e.g. numeric (1D, 2D, 3D, etc.), text, multiple-choice. Numeric fields in a context may be ranges of values rather than single values, e.g. a range of opening hours, or a 2D area represented by a circle or polygon.
  4. the context-aware application involves a changing current context; most often the current context relates to mobile objects, and the current context has a location field that is regularly changing. Normally the end-user receiving the retrieved information will be one of the objects from which the current context is set. (On the other hand the end-user could be someone in an office receiving information relating to the contexts of other objects such as moving vehicles, animals, people, etc.)
  5. the overall retrieval performed by an application may be a multi-stage operation, with, say, CAR being followed by a conventional IR activity that extracts those documents that relate to the user's current interests, as represented by the IR query.
  6. CAR may be a mixture of proactive and interactive activity. On the proactive side, a context attached to a document might, when it matches the current context, cause the document to be sent to the user; on the interactive side, a query might be constructed from the current context, and this may be matched against the documents in a collection. Thus CAR can be a mixture of IF (proactive) and IR (interactive).
  7. CAR applications may be nearly continuous: for example retrieval activity may happen every second, as the current context changes.
  8. CAR applications need fast retrieval: even if the application is not continuous, if the current context is changing, then, if newly retrieved information arrives so slowly that it relates to a long-past context, it may be useless.
  9. CAR will often be a background activity, going on while the user is pursuing some other task. In such cases, retrieved information will intrude on the user's foreground task: the need for high precision is therefore paramount.
  10. CAR applications may be highly distributed.
  11. in many applications the document collection will be changing dynamically, e.g. the documents may relate to traffic problems. Sometimes whole documents will be added or deleted; sometimes the content, and in particular the context, of a document may change, e.g. a traffic problem may relate to an increasingly large area. However some CAR systems may treat a change in document content as the deletion of the original document followed by the addition of a new one.
  12. often the results of retrieval will be presented on a small hand-held device carried by the user; the effect of this is largely in the user interface, but it may affect the retrieval approach (e.g. precision is weighted even higher).

Some relevant papers

1.
M. Korkea-aho, Context-Aware application survey, Helsinki Univ. of Technology, 2000.
2.
Context-awareness: some compelling applications, P.J. Brown, W. Burleson, M. Lamming, O-W Rahlff, G. Ramano, J. Scholtz, D. Snowdon. submitted for publication.
3.
todo: more