This paper proposes a continous space text model based on Gaussian processes. Introducing latent coordinates of words over which the Gaussian process is defined, we can encode word correlations directly and lead to a model that performs better than mixture models. Our model would serve as a foundation of more complex text models and also as a statistical visualization of texts.

Recently we have proposed Gaussian mixtures as a local statistical model to synthesize artificial textures. We describe the statistical dependence of pixels of a movable window by multivariate Gaussian mixture of product components. The mixture components correspond to different variants of image patches as they appear in the window. In this sense they can be used to identify different segments...