16.2 Why Overview Maps?

Overview maps are one of the best tools for the orientation and navigation in hypermedia documents [Utt89][Fei88][Tri88]. Authors disagree, however, whether these maps can be produced automatically. While there are systems that offer automatically generated overview maps including Intermedia and Hyper-G, Bernstein claims in [Ber88] that it may be impossible to generate satisfactory and comprehensible maps without understanding the contents of the documents.

Bernstein's system does not automatically generate overview maps, but instead shifts the burden of cartography entirely to the author. Bernstein lists two main reasons for the difficulty of the automatic generation of overview maps:

• A suitable graph structure for a two dimensional projection of the generally multidimensional hypertext network is hard to find. In particular this task is impossible without knowledge about the contents of the document.
• It is impossible to draw more than at most twenty node-icons on a screen. For every document that contains more nodes, many nodes must be suppressed to keep the screen from becoming a meaningless tangle. This requirement demands not only knowledge about the contents of the document, but also about the intentions and goals of the reader.

We agree with Bernstein in the reasons why it is so difficult to automatically generate an overview map. But instead of shifting this task to the hypertext author we rather try to solve the underlying problems. Our approach is different from the Intermedia and Hyper-G solution in that we do not use directly the actual linking structure of the document. Instead, we analyze the contents of the nodes and of the hyperdocument by using automatic indexing and clustering techniques as described in chapter 1. Based on this knowledge, we compute a structure graph for the overview map. This structure graph has nothing to do with the actual node-link structure of the hyperdocument.

We thus address the problem ofautomatic generation of structural overview maps of large collections of loosely structured data and the related problem of quickly presenting an overview of retrieved data. We will first depict an ideal solution. We then address step-by-step the problems raised by the ideal solution and describe a possible implementation. We later extend our system to include hierarchies of CYBERMAPs for very large node sets. We also present cybertrees, a minimum-spanning-tree-based algorithms for the automatic generation of tree shaped overview maps.