On-demand publishing is used to provide readers with fresh information, as well as to customize publications for different user profiles. When customized on-demand publishing is taken to the extreme, each customer receives an individually designed edition of the publication. In practice, this degree of customization can only be achieved when the publication process is nearly automatic. Currently, the part of the process that most readily may be automated is the pagination. As a multitude of tools and standards for publishing on the electronic media have emerged in recent years, automatic pagination has become an interesting option for a variety of situations.
Automatic pagination may be useful anywhere where pagination is considered useful but too expensive or too slow by traditional means. The presence of the following factors might therefore suggest using an application carrying out automatic pagination:
Availability of some kind of automatic pagination might enable small publishers or even individual people to produce high-quality publications. This may be the case for example with newspapers delivered by fax to a small group of users, where the contents are restricted excerpts from traditional newspapers. The situation also arises in the Internet, where anyone who wishes is able to become an information provider with very small investment.
Sometimes there is a need for fast delivery of information to the customers, and in the extreme case even a "quick and dirty" interactive pagination would be too slow. Despite that pagination could then be skipped altogether and the items could just be put on page one after another, visuality may nevertheless be considered important.
Publications where the information content heavily varies for different users or user groups, must accordingly be visualized individually. This situation may arrive with customized newspaper, database searches etc.
In some publications, the information content presented may be identical for everybody, but the layout requirements (such as page size) vary between users. This situation arises often with publishing on the electronic media, where screen sizes may vary considerably. And not only the screen sizes: the user may wish to view several different publications or documents simultaneously on the same screen. A publication with dimensions that can be adapted at time of view would be useful. Also the medium of the publication may vary: sometimes the user may prefer to read from the screen, at other times from a print.
The production of a page layout consists of two main tasks: Selecting the material to be paginated, and designing a page layout for the material. These tasks may be carried out sequentially, or they may be mixed by letting the optimization of the visual outlook affect the selection of material. For example, filler material may be used to fill spots that would otherwise remain empty, or an article version may be replaced with another if that eventually seems to fit better to the page. These tasks are affected by the different goals that can be associated with pagination (see Table 1 in section 3.1). The balance between partially conflicting goals such as minimization of empty space and maximization of high-priority material on page must be decided by the publisher.
The content of the generalized publication may be obtained from a large collection of material, which may be viewed as a kind of database. The database might, for example, consist of all the abstracts of scientific articles obtainable from the Internet via WWW searches (see section 2.3.1) or all the Usenet News articles of the group sci.med, or all the stories written by the editors of the electronic newspaper that one orders.
From the database, one probably would want to restrict the amount of material eventually viewed. This might be preferred in one of several ways of classifying or scoring text chunks based on rules, keywords, hierarchy of subject classes, different properties of the pieces of material such as how recent they are, who wrote them, subject etc. Data mining techniques or methods such as classification by artificial neural networks algorithms might also be used for this task. There exist learning algorithms that recognize the reader's reading pattern from previous experience and match it against yet unseen material (see for example the adaptive scoring of Ding Gnus by Ingebritsen, 1994). This kind of categorization might also be carried out by associative neural networks such as SOM [Kohonen, 1995].