In this paper authors carry out the research of three different topic models in the context of analyzing different large scale twitter ad hoc discussions, obtained from social network on different social and political events. An experiment is being conducted to test the effectiveness of models in analyzing three known discussions: the riots in Biryulyovo (Russia), Ferguson unrest (USA), Charlie Hebdo shooting (France). The results of the experiment show that the BTM topic model in terms of both Umass and Npmi, yielded the best results on all discussions in comparison to the baseline LDA model and another model for short imbalanced texts - WNTM.