COWORDS: a probabilistic model for text visualization

Abstract

This thesis introduces COWORDS, a new stochastic algorithm to create multiple word clouds, one for each document. The shared words in multiple documents are placed in the same position in all clouds. Similar documents produce similar and compact clouds, making easier to compare and interpret simultaneously several word clouds. The algorithm is based on a probability distribution in which the most probable congurations are those with a desirable visual aspect, such as a low value for the total distance between the words in all clouds. The visual aspect and the probabilistic model are guided by three principles: (i) tightness: it requires that the returned congurations should have all clouds with a minimum empty space amount between the words; (ii) overlapping: the words in each cloud must have no overlap; (iii) position: the words must be in the same spatial location in each cloud where they appear. The word congurations that do not follow these principles have a low probability of being observed. We built a Metropolis-Hastings algorithm, a special case of a Markov Chain Monte Carlo (MCMC) simulation method, to sample from the proposed clouds probability distribution. Our algorithm can easily incorporate additional constraints besides requiring the same position of the words in the dierent clouds. In addition, an extension of COWORDS is proposed. This extension allows the COWORDS algorithm to generate temporal word clouds preserving the semantic position of the words across all clouds. This new feature keeps the three main principles of COWORDS and adds one more: semantic: the words semantically correlated must be close to each other in all word clouds. Several simulation studies as well as case studies are conducted to evaluate and demonstrate the eectiveness of the COWORDS algorithm.