Brandes and Corman describe a method of visualizing a series of changes in a graph (an evolving network) as a stack of layers, one for each time step. Each vertex is thus a column at a fixed point through the layers in which it persists. A challenge arises in creating a layout for the graph that works for all layers under this constraint. It is necessary to assume that the graph is known in advance for all time steps--that is, it is not evolving in real-time. An iterative incremental algorithm that uses local scaling is presented for this.

The authors demonstrate their method on graphs that are said to represent the sentences of a conversation unfolding in time, and the paper is as much about the underlying theory of centering resonance analysis as it is about visualization. The method is essentially to regard noun and adjective word-types as vertices, and proximity of instances of word-types in a text as edges.

For example, a word-type that occurs in two sentences would be connected to its context in each of them. This allows a graph-theoretic definition of which words are most central to the discourse, a property that is represented in the visualization by the thickness of the column at each layer. Color is also used. However, the illustrations in the paper are undersized gray-on-gray halftones, in which it is very difficult to verify that the example visualizations have the properties that the authors attribute to them.

Using explicit or implicit graphs or chains that represent the repetition of words or the occurrence of related words is a familiar idea in computational linguistics and discourse analysis, and has applications in natural language processing (for example, Hoey [1], Morris and Hirst [2], and Hearst [3]). However, there is no discussion of how the present work differs from its predecessors. In fact, the paper doesn’t cite a single reference in the field, even though the authors claim that their method “uses computational linguistics.”