Social Filtering of News Using the Konstanz Hypertext System


M. Eisenmann and M. Rittberger

Department of Information Science
University of Konstanz
D-78457 Konstanz


The Konstanz Hypertext System has been expanded to process news. With the Konstanz Hypertext System users can read, arrange and classify news and send reports to news groups. With the aid of the knowledge base integrated into the Konstanz Hypertext System a social filter was added to the system. For the social filter the concept of the group was introduced, which is of central importance for the filtering of individual news articles. The filter developed was compared with other systems which provide filtering for news or e-mail.


Filter systems are usually designed for a specific use, since optimal efficiency can only be achieved when not only the special features of information sources, but also the particular characteristics of the user model and a stable and specific information need are drawn upon [Loeb 1992, Oard 1997].
Initial work on filtering mail or news articles was done in the context of the Information Lens system [Malone et al. 1987], HyperNews [Andersen et al. 1989], Mafia [Lutz 1990] and Iscreen [Pollock 1987]. Due to progress in network technique and the availability of networks, especially the Internet, other systems for filtering news and e-mail have been created, e.g., Tapestry [Goldberg et al. 1992, Terry 1993], Sift [Yan & Garcia-Molina 1995], Rama [Binkley & Young 1995], Similog [Horn 1995], Borges [Smeaton 1996], Infos [Mock & Rao Vermuri 1997], GrassRoot [Kamiya et al. 1996], GroupLens [Resnick et al. 1994, Konstan et al. 1997], Newsweeder [Lang 1997], spynews [Lueg 1998], and a neural network approach [Jennings & Higuchi 1993]. Network technique helped pave the way for a new generation of tools, the intelligent agents [Wooldridge & Jennings 1995, Maes 1994], which can solve problems in networks autonomously and cooperatively, even for the filtering of news (Infoscope [Fischer & Stevens 1991, Stevens 1992] or Ringo [Shardanand & Maes 1995]). Increasingly filters are also used in recommendation systems. They employ the judgements of users to make recommendations for other users, for example, when reading news with the Phoaks system [Terveen et al. 1997a, Terveen et al. 1997B].

In constructing filter systems not only filtering technology and user models are necessary [Paepcke et al. 1991], but a variety of criteria must be taken into account (domains, interaction, representation, user models and the filter itself) which have a considerable influence on the efficiency and usefulness of a filter system. We wish to discuss these criteria in the context of the named filter-based systems which process mail or news articles and finally present our own solution based on the Konstanz Hypertext System.

Domain: Domains are used to set the contents of the knowledge base which function in the filtering procedure. A domain is characterized by the texts, thus the articles of the news group, the participants of the news group with their aims, plans and communication behavior and by the users of the filters, who with their special aims and interests affect the filter function. The texts are determined by the types of document available, their formal preparation, the quantity of documents and the time frames in which new documents appear in the system. Thus, for example, it is of great importance whether the filter system draws on

  1. semi-structured documents (Information Lens, Iscreen, GroupLens, Tapestry, Infoscope, Infos), as is usual, e.g., with news,
  2. text documents, respectively only the text base (HyperNews, Borges, Newsweeeder),
  3. various document types (Rama, Phoaks, Sift).
With filter systems the text is particularly focused on, since it usually describes the passages of a document with relevant contents.
Often there is a different degree and specific type of inter-networking among documents. Thus just on the basis of these differences e-mails, mailing lists and news groups must be handled differently. The structuring or classifying of e-mails or news groups can be based on formal statements, for example the repeated occurrence of authors' names in e-mails. Such formal structures can be easily drawn upon and are therefore directly available to a filter system. With Rama, directory structures are used for semantic statements linking individual documents and employed during filtering. With HyperNews the hierarchical structures of news are used in the filtering process. However, these formal structures must be tested with the contents of the text parts, since formal structures do not necessarily have to correspond to the intentions of authors in writing e-mails, and incorrect allocations can thereby arise in filtering.

Interaction: The acceptance of filter-based systems depends essentially on appropriate user interaction [Binkley & Young 1995, Malone et al., 1987, Pollock 1987]. Distinctions can be made concerning the manner in which documents are filtered and made available to the user by means of an adequate user profile and the way the user interacts with the filter itself [Lueg 1998]. The applicability of a filter is also important in the interaction. Usually filters should protect the user from being deluged by an excessively large quantity of information and should thus interact with other systems. Therefore one can distinguish between filter systems:

Representation: Semantic representation is done according to different criteria with different systems, e.g., with vector based (Rama, Borges, HyperNews, Sift) or probabilistic methods, semantic classification (Mafia, Infos) or the use of templates (Iscreen) for the formal analysis of contents. Only with Mafia is there a semantic analysis of the contents. The other systems attempt to make statements about a text or the similarity of texts indirectly by means of the formal characteristics of the text. If only formal characteristics are used in a content analysis, the chances of success are proportional to the qualitatively high-value characteristics of a document. Documents from the Internet in particular frequently possess such characteristics, as for example authors, links to other documents, the address of the server, an access path, etc. Such a knowledge-based representation occurs, for example, with the Mafia, Infos and Iscreen systems. However, they do not offer a complete representation of the domains, since, for example, authors are not modeled in Iscreen.
Presentation: According to [Belkin & Croft 1992] filters are used by the recipients of information. Filters are therefore usually oriented to already existing information flows, formats and user interfaces. In some systems information-rich titles or other structuring attributes such as ratings (GroupLens or Phoaks) are attached for presentation. These ratings usually involve the sorting of articles and the presentation of a sequence based on the relevance of the documents. Further sorting attributes such as the creation of additional virtual news groups in Infoscope contribute to orientation. For presentation additional formatting and indexing to the individual articles are often added.

Naturally with newer works there are graphic interaction tools which permit access to information and the visualization of contexts. Negatively filtered results are either discarded or can be accessed through an archive.

User Model: Filters are employed for continuing interest in a question, which is one of the chief distinguishing attributes of filtering for information retrieval. If one views this interest as the most important attribute of a user model, mechanisms must be included in the filter which make it possible to create and modify a user model. In particular it must be easy to make changes while using the filter, since the variability of users' interests over a longer period of time inevitably requires that filters be highly flexible [Mock & Rao Vermuri 1997]. Changes are based on judgements by users of the quality of the filtered documents and on a shift in users' interests influenced by the documents viewed and rejected [Jennings & Higuchi 1993]. Besides the modeling of user demands for filter results, additional assumptions can also be made which, for example, are defined in stereotypes (Mafia) to help the system with unknown information about the user [Shapira et al.1997]. In the systems discussed the same formats are used for the representation of the user model as for the representation of the other areas of the domain. Common to all systems is that the user model is created by the users themselves. If their interests or the quality of the results change, changes or corrections in the user profile become necessary which the users themselves must make.

Filters: [Oard 1997] discusses filters under terminolgical aspects and shows the relationship to 'Routing', 'Selective Dissemination of Information (SDI)', 'current awareness' and 'Recommendation'. On the basis of evaluations [Malone et al. 1987, Goldberg et al. 1992] distinguished three types of filter which are employed by users in working with different types of information:

  1. With cognitive or content based filters the filter criteria are set on the basis of the attributes of a document or the content of a text. This includes the employment of index terms, the date of production or the identification of a specific topic. A complete semantic analysis (Mafia) or even only a partial analysis, such as of the title (e.g., Tapestry) is thereby seldom made. Infos combines a content filter with a collaborative filter which, however, plays only a subordinate role in the system's relevance decision.
  2. With economical filters filtering depends on the length of the document, thus the time needed by the user to acquire and understand the document.
  3. Social or collaborative filters are based on structures connecting communication partners. With collaborative filtering systems, sometimes also mentioned as recommender systems [Resnick & Varian 1997], the filtering results on an user profile of one person which is used by others. It is necessary for this end that the persons who use the profile have information about the filtering behavior of the other. The platform for internet content selection (PICS) [Resnick 1997] offers a method to label software, information, etc. with information about the authors of the labeled piece. The problem of trusting persons, not software or information services is still relevant with PICS, but labeling gives a better basis for trust decisions [Kuhlen 1999]. 
    Collaborative filtering thus presupposes knowledge about the organizational structure and its employment in social filtering. Such structures are, for example, employed in Grassroots to filter and categorize information. This type of filtering is found especially in the application domain of news, since interesting information can be acquired from the high level of structuring and the many linkages of news (e.g., Tapestry, GroupLens, Phoaks). A mixed form of collaborative and cognitive filtering describes the ordering of news articles by so-called 'Collectors' [Kamiya et al. 1996]. In them the news articles are arranged on the basis of content or respectively group-specific criteria and hierarchies of these criteria.
Taking into account the named criteria, the knowledge base in the Konstanz Hypertext System (KHS) was used to model a filtering process which accesses the semi-structured data of news groups. The procedure of the filter is thereby set by the communication behavior of the user in the news group. In the following section a description will be given of the domain of the news group, the communication model underlying it and the representation of news in the KHS. After a brief introduction to the KHS the conception of the KHS news filter will be explained, and our user model will be described. Finally, concrete modeling will be described.

News group and communication behavior

In classical news readers, such as, e.g., Netscape, the user has various interaction possibilities in his domain. Three user types can be distinguished:
In order to do justice to the interaction forms of the various user types in the news group domain the filter model of [Belkin & Croft 1992] was modified (Fig. 1). In [Belkin & Croft 1992] a filter model is proposed which describes two user groups. One consists of authors who create unstructured or weakly structured news articles in their domain which are then presented and distributed on the Internet. Another consists of participants, whose interest in the news group is represented by its user profile. To modify the filtering results users can either correct the user profile (1) or in a further feedback the interests of the user can be influenced by the evaluation of the filtering results (2). In this model the simultaneous occurrence of author and reader as a single person is not taken into account. This attribute can also have considerable influence on the user profile. We therefore propose an expansion (3) of the model in order that the contribution of the author can be directly employed in constructing a user profile. If one expands this conclusion from individual users to all authors in a news group, it is possible to maintain the user profile of each individual author in the news group. However, the quality of the profile thereby depends heavily on the activity of the individual author.
Besides the activities of the news group participants, the formal structure of the news group can be employed to derive user profiles. As with e-mails individual articles are formally designated by author and subject. In addition, news groups are also characterized by the tree-like structure of the threads which result from the mutual references in the articles. Different topics can thus be represented by different subjects or different sub-trees, whereby combinations are also possible. If, e.g., an author changes the subject within a thread, it breaks down into two sub-trees. The dynamic behavior of the news group is characterized by the temporarily limited storage of the news article in a specific period of time. Besides the technical problem of the storage and presentation of the filtering results, it is thereby of particular interest that the user profile can also change dynamically.

Figure 1: Filtering model modified for news groups based on [Belkin & Croft 1992].

Since the user can freely choose the thread to which he sends his article and which topic he specifies, the structure of the news group is strongly affected by the discussion behavior of the users. The formal hierarchy does not necessarily reproduce the actual discussion structure. Thus, e.g., sometimes the topic of the contribution changes but the user does not adjust this in the subject field, or due to an error the author specifies a new subject. Due to the large number of participants, who are usually unknown to the author, relatively great social control is present, so that discussion behavior can also be employed to derive a user profile.

The structural attributes of the news group presented here permit different ordering principles, with which the domain can be analyzed. Since many news readers use the ordering principles for presentation, filtering results must be made accessible to users in a similar way, in order to achieve the necessary user acceptance. The following ordering principles can be employed:

Attention should be paid in particular to currency and topicality. With classical news readers such as Netscape the expression of currency and topicality is limited to indicating whether a user has already read a contribution or not. Articles which have been read can thereby also be marked as unread. If the user has not read news group articles for some time, he receives all the new articles along with the unread marked articles. Consequently in this context currency and topicality does not necessarily mean current and topical for the user.
The possibility to combine the above-named ordering criteria [Kamiya et al. 1996] in order to further navigate in the structure of the thread is in itself a simple form of filter, such as has already been implemented for the processing of mail in the Konstanz Hypertext System [Hammwöhner & Rittberger 1997].

The Konstanz Hypertext System

The Konstanz Hypertext System (KHS) meets the current criteria for openness in hypertext systems [Hammwöhner 1997, p. 99]. It can be employed in a variety of ways and is adaptable to quite different applications [Hammwöhner & Rittberger 1997]. This flexibility is provided by a variety of attributes which are integrated into the architecture of the system.
The unifying framework of the KHS is supplied by a generic, application independent hypertext-model comprising a structure model which describes the structure of well-formed hypertexts and an interaction model which defines generic interaction styles. Both the structure model and the interaction model can be refined to suit the needs of special applications or individual users. Because the interaction model is of little consequence in our context (a description of the interaction model is provided in [Rittberger 1994]), the structure model will be described in some detail (for more detail see [Hammwöhner & Rittberger 1997]). The KHS is implemented in Smalltalk and available on several plattforms.

By typing hypertext objects, the KHS allows users to structure a hypertext in the way most suitable for their current application. KHS units are arranged in a polyhierarchy of composite units which allows users to structure information in a domain specific way, e.g. all information about a specific topic being researched can be assembled and organized under one composite node. The system also distinguishes between composite units and terminal units.

Besides these basic attributes of the KHS, two aspects of the KHS play an essential role in our context. First, there is the integration of external data sources and their further processing using the KHS. In [Hammwöhner & Rittberger 1997] the framework conditions for the processing of external information resources are explained and examples are offered of the further processing of data for information resource selection, organization and communication with e-mail and the management of electronic journals. Second, a knowledge base is integrated into the KHS. The knowledge-based component makes available three mutually enhancing, different forms of representation: frames, access paths and rules. For the representation of declarative knowledge a frame formalism [Reimer 1991] was integrated into the KHS and a system of production rules [Puppe 1991] was employed to formulate procedural knowledge [Zink 1995]. In addition, so-called access paths offer a simple, unified access from procedural to declarative knowledge. The knowledge base is built into the KHS by the use of typified links and nodes, so that all formally represented knowledge objects are present in the KHS as hypertext objects. Besides static knowledge, which is primarily represented by frames, tasks defined by rules can be employed for modeling in the KHS. In order to check the applicability of the knowledge-based components, an updating of the integrated information services in the KHS was planned. One of the most widespread applications of the Internet, beside e-mail, Telnet and WWW, which are already available in the KHS, is news. Since large quantities of data must be managed in news, the integration of news in the KHS was combined with the filtering of news articles with the aid of the knowledge base.

The integration of news into the KHS depends strongly on the already present functionality for the processing of e-mail. News articles can be integrated into the poly-hierarchical structure of a KHS hypertext, and articles by users can be sent to the news group. The individual threads of a news group are deposited in separate folders. The relationships between the individual articles of a thread are represented by links. Furthermore, a classification can be made based on attributes from either the subject field or the sender field. Thus, for example, classifications can be made based on specific authors or on the basis of specific concepts in the subject field. In addition, linkages can be set up between news articles based on statistical similarities of the texts using the vector space model. Using the cassation available in the KHS, the expiration date of the news article can be specified. Various models are usable for this:

Here as well the cassation can be combined with individual folders so that, for example, news articles by a specific author can be stored for a long time, but specific topics only for a short time.

Concept of the KHS Filter

The various concepts available in the knowledge base of the Konstanz Hypertext System are employed to represent the knowledge needed for the filtering procedure. Each article of a news group is represented by a frame. The individual fields of a news article are represented by slots of the frames. Access to the structure of the news group, i.e., to the links between the units, is made using access paths. A special role is played by the author and subject fields, which are represented as individual frames. In addition, the hypertext units of a KHS hypertext which contain the news article are entered into a slot. A frame is additionally created containing the information which represents the structure of the thread. A news reader is installed in the KHS which observes the activities of a news group and reads in new news articles. The frames for the representation of the news articles and the threads are automatically created while news articles are being read in. In order to guarantee the full functionality of the system and to steer the activities derived from the knowledge base, in the KHS news articles are represented as frames and as 'normal' hypertext units.


Figure 2: Conception of the KHS News Filter

Besides the directly understandable concepts of author, subject, thread and news article, the concept of the group is employed in modeling. Groups appear in news in different concepts:

Forum, folder and thread thus describe groups of objects which have something in common. These concepts of the knowledge base are assigned a number of attributes:
In Figure 2 a concept of news filter is represented which is based on the expanded model of [Belkin & Croft 1992]. In Basic Knowledge knowledge is stored which directly represents threads and news articles. Derived Knowledge is deduced from Basic Knowledge. It is only linked to the Basic Knowledge and is eliminated when news articles expire. It contains knowledge on the structure of the news group in addition to the information of the thread, thus, for example, which authors respond to the news articles of other authors.

Knowledge about users is deposited in three different user profiles. These profiles are very simply maintained and describe the interest or lack of interest of the users in individual threads, news articles, persons or groups:

News articles are deposited in the three folders of Relevance Unknown, Irrelevant and Relevant. The news articles from the Relevant and Relevance Unknown folders are presented to the user. Depending on his judgement, the news article is deposited in the folders User Relevant or User Irrelevant. As the user takes advantage of this possibility, certain knowledge can be inferred from the supposed knowledge in the three folders, Relevance Unknown, Irrelevant and Relevant. The news articles in the three folders are deleted and deposited in the two folders User Relevant and User Irrelevant. The irrelevant news articles are initially stored in the system, because due to the high degree of structuring of news a high dependency exists between individual news articles, and the case can occur where the user decides when navigating through a thread that he really does want to read an article rejected as irrelevant.
The knowledge derived from the relevance judgements of the user is deposited in Feedback Knowledge using the Feedback Rules. In order not to confuse the user by constantly changing the profile, the tuning of the modeled profiles which are used for filtering is left up to the user. He can himself change the entries in the Modeled Profile on the basis of information from the Feedback Profile.

The changes described in [Belkin & Croft 1992] in filtering processes through the correction of the user profile or the adaptation of goals and interests are covered in the KHS knowledge base by the Feedback Profile. Based on the observation of the news group itself through the Derived Knowledge one receives a user profile with the Derived Profile which is based on the users' activity and corresponds to the feedback (3) in Figure 2.

The filtering process itself is initiated by the user. In the first phase the filter can only access the Derived Profile, since no information is as of yet deposited in the Modeled Profile. The filtering results are presented to the user in the Hypertext Browser, the main window of the KHS. Thereby the user can select various entry points to the newsgroup articles. Besides entry through news articles, entire threads can be offered, insofar as they are relevant. Furthermore, the user can receive new, relevant news articles related to specific topics or authors.

Depending on the complete integration of the knowledge base in the total system, the user can not only navigate in the news articles and the structures underlying them, but can also employ the additionally created structures, such as the discussion fora, as structuring means in order to obtain further access possibilities.

Figure 3: Classification of the KHS Filter in comparison with other systems.

Classification of the KHS Filters and Discussion

Networked environments frequently also set conditions on a high degree of structuring of documents, i.e., that implicit (e.g., databases, news) or explicit linkages (links which are also formally represented) are present. In Figure 3 the degree of structuring of the systems discussed is related to the methods of interpreting documents. The interpretation describes the procedure for content analyses. We thereby distinguish between classical indexing procedures which, for example, work according to the vector space model and knowledge-based procedures. The representation forms of knowledge-based systems are either oriented according to formal perspectives to the structure of the text, or the contents of the texts themselves are represented. A large number of the systems considered thereby use classic procedures for inferring content (e.g., a vector space model) or work with collaborative filters which use formal, knowledge-based analyses. Only a very few systems analyze the texts of mail or news articles semantically with the assistance of a knowledge base. The high degree of linkage and structuring, especially with news, of the articles is only partially employed for the filtering process or re-offered in the user interfaces.
In the KHS formal characteristics of news groups and news articles are employed in order to process knowledge about the news groups and news articles. This knowledge is used for the filtering process. The classificatory instruments available in the KHS and indexing with the aid of vector space models are not yet taken into account in the filtering process. The high degree of linkage and structuring of news can be excellently represented with an elaborated hypertext system like the KHS. In addition, still further ordering and structuring attributes can be integrated to aid the user in navigating or make possible access to the articles of a news group. The following objects of the knowledge base are especially suitable for this:
An initial check of the filter function was made on the basis of two scenarios. At a fixed point in time the contents of the news group, which serves as a discussion group for sport divers, were loaded and filtered in the KHS. A total of 148 and 154 news articles were respectively evaluated by the filter and two test persons. It was found that the behavior of the filters achieved values acceptable to the active test person. Thus 69 of the 97 news articles which the system proposed were among the 83 documents classified as relevant by this user. With the non-active user the filter without the possibility of deriving concrete information from user behavior was found to be clearly less able to identify relevant news articles. Thus the non-active user judged 53 of the 148 articles as relevant, whereby the KHS suggested only 26 news articles to him, of which half were also classified by this user as relevant. [Oard 1997] mentiones the necessity of a broader number of users for evaluationg social filtering systems. To confirm the optimistic points of our comtemplation a more detailed evaluation has to be done. Nevertheless the abilities of KHS to organize an information space for a individuell person combined with the introduced feddback techniques and social filtering capabilities seems to be a promising approach.
This initial study shows that the social filter for an active participant in a news group is a suitable tool for making relevant news articles available to him. The non-active user benefits little from the social filter and in the KHS must go through the knowledge base, indexing or classification in order to find possible alternative access paths to news articles.


[Belkin & Croft 1992] N.J. Belkin and W.B. Croft. Information Filtering and Information Retrieval: Two Sides of the Same Coin? Communications of the ACM (35)12, 29-38, 1992
[Binkley & Young 1995] J. Binkley and L. Young. Rama: an architecture for Internet information filtering. Journal of Intelligent Information Systems: Integrating Artificial Intelligence and Database Technologies (5)2, 81-99,1995

[Fischer & Stevens 1991] G. Fischer and C. Stevens. Information access in complex, poorly structured information spaces. Proceedings of CHI'91 ACM: New York, NY, 63-70, 1991

[Goldberg et al. 1992] D. Goldberg, D. Nichols, B.M. Oki and D. Terry. Using collaborative filtering to weave an information tapestry. Communications of the ACM (35)12, 61-70, 1992

[Hammwöhner 1997] R. Hammwöhner. Offene Hypertextsysteme. Das Konstanzer Hypertextsystem (KHS) im wissenschaftlichen Kontext. Universitätsverlag Konstanz, No. 32 Schriften zur Informationswissenschaft,1997

[Hammwöhner & Rittberger 1997] R. Hammwöhner and M. Rittberger. Building application dependent hypertexts. Information Processing & Management (33)2, 243-254, 1997

[Horn 1995] W. Horn. Information Filtering von Usenet News. Proceedings Workshop"Multimediale Informations- und Kommunikationssysteme", 23-30, 1995

[Jennings & Higuchi 1993] A. Jennings / H. Higuchi. A user model neural network for a personal news service. In: User Modeling and User-Adapted Interaction (3)1, 1-25, 1993

[Kamiya et al. 1996] K. Kamiya, M. Roscheisen and T. Winograd. Grassroots: a system providing a uniform framework for communicating, structuring, sharing information, and organizing people. Conference: Fifth International World Wide Web Conference. Paris, France, 6-10 May 1996 Computer Networks and ISDN Systems (28)7-11, 1157-1174, 1996

[Kautz et al. 1997] H. Kautz, B. Selman and M. Shah. ReferralWeb: Combining Social Networks and Collaborative Filtering. Communications of the ACM (40)2, 63-65, 1997

[Konstan et al. 1997] J.A. Konstan, B.N. Miller, D. Maltz, J.L. Herlocker, L.R. Gordon and J. Riedl. GroupLens: applying collaborative filtering to Usenet news. Communications of the ACM (40)3, 77-87, 1997

[Kuhlen 1999] R. Kuhlen. Die Konsequenzen von Informationsassistenten. Was bedeutet informationelle Autonomie oder wie kann Vertrauen in elektronische Dienste in offenen Informationsmärkten gesichert werden?. to be published, Suhrkamp: Frankfurt a. Main, suhrkamp taschenbuch wissenschaft 1999

[Lang 1995] K. Lang. NewsWeeder: learning to filter netnews. In: Machine Learning. Proceedings of the Twelfth International Conference on Machine Learning, 331-339, Morgan Kaufmann Publishers :San Francisco, CA, 1995

[Loeb 1992] S. Loeb. Architecting personalized delivery of multimedia information. Communications of the ACM (35)12, 39-48, 1992

[Lueg 1998] C. Lueg. Supporting situated actions in high volume conversational data situations. In: Proceedings of the 1998 Conference on Human Factors in Computing Systems, CHI, 472-479, ACM: New York, NY, 1998

[Lutz et al. 1990] E. Lutz, H. von Kleist-Retzow and K. Hoernig. MAFIA - an active mail-filter-agent for an intelligent document processing support. ACM SIGOIS Bulletin (11)4, 16-32, 1990

[Maes 1994] P. Maes. Agents that reduce work and information overload. Communications of the ACM (37)7, 30-40,1994

[Malone et al. 1987] T.W. Malone, K.R. Grant, F.A. Turbak, S.A Brobst and M.D Cohen. Intelligent information-sharing systems. Communications of the ACM (30)5, 390-402, 1987

[Mock & Rao Vermuri 1997] K.J. Mock / V. Rao Vermuri. Information Filtering via hill climbing, WordNet, and index patterns. Information Processing & Management (33)5, 633-644,1997

[Oard 1997] D. Oard. The state of the art in text filtering. In: User Modeling and User-Adapted Interaction (7)3, 141-178, 1997

[Paepcke et al. 1991] A. Paepcke / D. Fafchamps / R. Jeffries. Designing contextual information filters. Oct. 1991, Hewlett-Packard Lab., Palo Alto, CA, USA, HPL-91-145, 1991

[Pollock 1988] S. Pollock. A Rule-Based Message Filtering System. ACM Transaction on Office Information Systems (6)3, 232-254, 1988

[Puppe 1991] F. Puppe. Wissensrepräsentation mit Regeln. In: P. Struß (ed.) Wissensrepräsentation, Oldenbourg Verlag, 123-130, 1991

[Reimer 1991] U. Reimer. Einführung in die Wissensrepräsentation. Netzartige und schema-basierte Repräsentationsformate. B.G. Teubner, Leitfäden der Angewandten Informatik, 1991

[Resnick et al. 1994] P. Resnick, N. Iacovou, M. Suchak, P. Bergstrom and J. Riedl. GroupLens: an open architecture for collaborative filtering of netnews. Transcending Boundaries, CSCW '94. Proceedings of the Conference on Computer Supported Cooperative Work 175-186 ACM: New York,1994

[Resnick 1997] P. Resnick. Filtering Information on the net. In: Scientific American 3, 1997

[Resnick & Varian 1997] P. Resnick / H.R. Varian. Recommender systems. In: Communications of the ACM (40)3, 56-58, 1997

[Rittberger 1994] M. Rittberger, R. Hammwöhner, R. Aßfalg and R. Kuhlen. A homogenous interaction platform for navigation and search in and from open hypertext systems. RIAO 94 Conference Proceedings. Intelligent multimedia information retrieval systems and management, 649-663,1994

[Shapira et al. 1997] B. Shapira / P. Shoval / U. Hanani. Stereotypes in information filtering systems. In: Information Processing & Management (33)3, 273-287, 1997

[Shardanand & Maes 1995] U. Shardanand and P. Maes. Social information filtering: Algorithms for automating "Word of Mouth". Proceeding of CHI'95, ACM: New York, NY, 210-217,1995

[Smeaton 1996] A.F. Smeaton. Using NLP resources in the BORGES information filtering tool. Electronic Library and Visual Information Research. ELVIRA 3 37-49 ASLIB: London, UK,1996

[Stevens 1992] C. Stevens. Automating the creation of information filters. Communications of the ACM (35)12, 48,1992

[Terry 1993] D.B. Terry. A tour through Tapestry. Conference on Organizational Computing Systems, 21-30, 1993

[Terveen et al. 1997a] L.G. Terveen, W.C. Hill, B. Amento, D. McDonald and J. Creter. Building task-specific interfaces to high volume conversational data. Proceedings of the 1997 Conference on Human Factors in Computing Systems, CHI, 226-233 ACM: New York,1997

[Terveen et al. 1997b] L. Terveen, W. Hill, B. Amento, D. McDonald and J. Creter. PHOAKS: a system for sharing recommendations. Communications of the ACM (40)3, 59-62, 1997

[Wooldridge & Jennings 1995] M. Wooldridge and N.R. Jennings. Intelligent Agents: Theory and Practice. The Knowledge Engineering Review (10)2, 115-152, 1995

[Yan & Garcia-Molina] T.W. Yan and H. Garcia-Molina. SIFT-a tool for wide-area information dissemination. Proceedings of the 1995 USENIX Technical Conference Berkeley, CA, USA: USENIX Assoc, 177-186, 1995

[Zink 1995] V. Zink. Die Wissensbasis im Konstanzer Hypertextsystem (KHS). Informationswissenschaft, Universität Konstanz, Report, 77-95, 1995