Navigating within news collections using tag-flakes

The use of tag clouds is common for presenting frequently occurring tags or keywords in a collection to the users. Most visualizations of tag clouds vary the sizes of the fonts to differentiate important tags from others. This, however, is sufficient neither to help the user explore and discover relationships between tags in a collection, nor to help track the changes in these relationships across time frames in dynamic collections. In this paper, we propose an alternative ". contextual-layout" method, tag-flakes, for presenting tags or keywords that are associated with dynamically evolving textual content, like news streams. A TMine algorithm first maps tags onto a latent semantic space. However, instead of using this latent semantic space to simply cluster and index the documents (as commonly done in many existing schemes), TMine analyzes the relationships between tags in this semantic space and the resulting tag cloud is condensed into a hierarchy (or a tag-flake) in a way that captures contextual relationships between tags: descendant terms in the hierarchy occur within the context defined by the ancestor terms. This provides a mechanism for navigation within the tag space as well as for the contextual organization of the text documents. We use TMine in developing the tagFlake visualization system, which relies on TMine for organizing tags extracted from news collections in a hierarchical manner and supports navigation within the collection through these contextually laid-out tag clouds. tagFlake also helps users track topic developments and changes in the context in which certain keywords are used. Experimental evaluation results show the effectiveness of the proposed TMine method in capturing the semantic structures of collections.