Iterative visual clustering for unstructured text mining

This paper proposes the iterative visual clustering (IVC) on unstructured text sequences to form and evaluate keyword clusters, based on which users can use visual analysis, domain knowledge to discover knowledge in the text. The text sequence data are broken down into a list representative keywords after textual evaluation, and the keywords are then grouped to form keyword clusters via an iterative stochastic process and are visualized as distributions over the time lines. The visual evaluation model provides shape evaluations as quantitative tools and users' interactions as qualitative tools to visually investigate the trends, patterns represented by the keyword clusters' distributions. The keyword clustering model, guided by the feedback of visual evaluations, step-wisely enumerates newer generations of keyword clusters and their patterns, therefore narrows down the search space. Then the proposed IVC is applied onto nursing narratives and is able to identify interesting keyword clusters implying hidden knowledge regarding to the working patterns and environment of registered nurses. The loop of producing next generation of keyword clusters in IVC is driven and controlled by users' perception, domain knowledge and interactions, and it is also guided by a stochastic search model. So both semantic and distribution features enable IVC to have significant applications as a text mining tool, on many other data sets, such as biomedical literatures.