A notable rise in the amounts of data collected, which were made available to the public, is witnessed. This allowed the emergence of many research problems among which extracting knowledge from short texts and their different related challenges. In this thesis, we elaborate new approaches to enhance short text clustering results obtained through the use of mixture models. 
We deployed the collapsed Gibbs sampling algorithm previously used with the Dirichlet Multinomial mixture model on our proposed statistical models. In particular, we proposed the collapsed Gibbs sampling generalized Dirichlet Multinomial (CGSGDM) and the collapsed Gibbs sampling Beta-Liouville Multinomial (CGSBLM) mixture models to cope with the challenges that come with short texts. We demonstrate the efficiency of our proposed approaches on the Google News corpora. We compared the experimental results with related works that made use of the Dirichlet distribution as a prior.
Finally, we scaled our work to use infinite mixture models namely collapsed Gibbs sampling infinite generalized Dirichlet Multinomial mixture model (CGSIGDMM) and collapsed Gibbs sampling infinite Beta-Liouville Multinomial mixture model (CGSIBLMM). We also evaluate our proposed approaches on the Tweet dataset additionally to the previously used Google News dataset. An improvement of the work is also proposed through an online clustering process demonstrating good performance on the same used datasets. A final application is presented to assess the robustness of the proposed framework in the presence of outliers.