Abstract—Centroid terms are comfortable instruments to represent texts, compare them semantically and to even (hierarchically) cluster sets of documents using them. Their determination depends on their topical and conceptual context, i.e. the dynamically changing knowledge of a user represented by the co-occurrence graph. Herein, important properties of centroids as well as their applicability for tasks in natural language processing and text mining shall be discussed and their use justified by a set of experiments. Based on the obtained results, a new approach to detect fine-grained similarities between text documents is derived.

 

Index Terms—centroid term, co-occurrence graph, document similarity, text processing

Cite: Mario Kubek, Thomas Böhme, and Herwig Unger, "Empiric Experiments with Text Representing Centroids," Lecture Notes on Information Theory, Vol. 5, No. 1, pp. 23-28, June 2017. doi: 10.18178/lnit.5.1.23-28