Affiliations: Centro de Estudios Superiores Felipe II, Universidad
Complutense, Aranjuez, Madrid, Spain | Departamento de Sistemas Informáticos y
Programación, Universidad Complutense Madrid, Spain. Tel.: +34
913947641; Fax: +34 913947529; E-mail: [email protected]
Abstract: The designer of an information dissemination system based on user
preferences stated as user models is currently faced with three basic design
decisions: whether to use categories, keywords – or both – to
enable the user to specify his preferences, whether to use a static long-term
model or a dynamic short-term model to register those preferences, and what
method to use to provide summaries of the available documents without losing
information that may be significant to a particular user even if it would not
be considered as such in general terms. Current systems tend to provide one
specific choice – either taken at design time by the developer or offered as
mutually exclusive alternatives to the user. However, most of the options have
relative merits. An efficient way of combining the various solutions would
allow users to select in each case the combination of alternatives better
suited for their needs. In this paper we defend the use of a combined approach
that integrates: an enriched user-model that the user can customise to capture
his long-term interests either in terms of categories (newspaper sections) or
keywords, a personalised summarization facility to maximise the density of
relevance of sent selections, and a tailored relevance feedback mechanism that
captures short-term interests as featured in a user's acceptance or rejection
of the news items received. Controlled experiments were carried out with a
group of users and satisfactory results were obtained, providing material for
further development. The experimental results suggest that categories and
keywords can be fruitfully combined to express user interests, and that
personalised summaries perform better than generic summaries at least in terms
of identifying documents that satisfy user preferences.
Keywords: Summarization, relevance feedback, personalisation, user modeling, text classification
