In contrast to IR, IF has only recently started to attract attention. Based on a survey of information sharing in organizations, three approaches can be identified [29]. Depending on the manner in which documents are selected for the user, filtering systems can be classified as cognitive, social or economic. Cognitive systems choose documents based on the characteristics of their contents. Social systems select documents based on the recommendations and annotations of other users. Economic systems select documents based on some computation of cost-benefit to the user and through some pricing mechanisms.
A variety of approaches have been used to get at the semantic contents of the documents. Oval [29] is an example of a system which uses a keyword-based approach to match user defined rules to incoming documents. Foltz [10] demonstrates the use of LSI for information filtering and evaluates it for filtering Netnews articles. [11] performs a similar experiment for the domain of technical reports. INFOSCOPE [9] consists of rule-based agents which observe usage patterns and make suggestions to the user. The agents monitor the contents of the messages that are deemed interesting or uninteresting, make statistical correlations and suggest changes to the user.
Social systems typically select documents based on the ratings that other users assign to them. Users collaborate to help each other filter documents. Eager readers would be the first ones to read incoming articles and provide their endorsements, which will be used by passive readers to filter articles, as in GoodNews [45]. The selection could either depend on personal criteria (e.g. endorsement by a friend), or on aggregate criteria (e.g. endorsements by more than half the group members). Tapestry [13] is an example of a collaborative environment that accepts information from many sources, allows detailed endorsements and defines a query language to access endorsed articles.
The cognitive and social approaches are both just as valid for selecting documents. The difference lies in the fact that depending on the application area, one is likely to be more valuable than the other. If information is being gathered for keeping up to date with a certain community, social filtering is the way to choose documents. However, if information is being gathered based on the topic, independent of who the other users of the information are, then cognitive systems are more appropriate. Of course, sometimes a hybrid approach is necessary.
Commercial filtering services have recently entered the market. ``First!'' is a personalized news clipping service provided by Individual, Inc. The user profile is acquired by talking to the users and having them fill out templates. Personalized news is delivered frequently through facsimile or email messages. The filtering is done by the Smart system [42], augmented by human supervision.