There has been major progress in recent years in statistical model-based pattern recognition, data mining and knowledge discovery. In particular, generative models are widely used and are very reliable in terms of overall performance. Success of these models hinges on their ability to construct a representation which captures the underlying statistical distribution of data. In this article we focus on count data modeling. Indeed, this kind of data is naturally generated in many contexts and in different application domains. Usually, models based on the multinomial assumption are used in this case which may have several shortcomings especially in the case of high-dimensional sparse data. We propose then a principled approach to smooth multinomials using a mixture of Beta-Liouville distributions which is learned to reflect and model prior beliefs about multinomial parameters. Via both theoretical interpretations and experimental validations, we argue that the proposed smoothing model is general and flexible enough to allow accurate representation of count data.