Predicting topic shift locations in design histories

Abstract

Knowing the locations of weak-to-strong topic shifts in a design history enables hierarchical segmentation of that history. The segmentation can be the basis of hierarchical visualization, that is, semantic zoom, and more, such as visualization of co-location, co-citation, and density of linking. This research shows that a fine-grained, sub-topical linkograph of a design conversation can be used to identify the locations of topic shifts in that conversation. A design conversation spanning 11 design meetings was captured; deictic (pointing-like) references were simulated by performing a sentence-by-sentence-level linkograph analysis of the conversation (that is, the conversation was not subjectively aggregated to topical segments prior to analysis); an algorithm used the linkograph to predict the locations of topic shifts; and the linkograph-predicted topic shift locations were compared with expected topic shift locations for the same conversation. The expected topic shift locations were defined as the heads of reference series (references to transcript units) that were made in a detailed report about the meetings. The model performed well (63–80 %) on large reference series (quartiles three and four) and poorly on small reference series (quartiles one and two, and singular references). Future research will work with linkographs that are automatically constructed by both text-based and graphic design systems and aim to develop a framework that can adapt to individual histories.