In this work, each text description will be represented with the posterior Dirichlet random variable θ which represents the topic distribution over k latent topics. However, it is not computationally feasible to learn the model directly from a given database, because it requires computing p(w|α, β) which includes intractable integral operations.