Over the past few years the sentiment analysis task of users' posts in social networks has become very popular among researchers. In this paper, authors present and describe the developed multi-lingual knowledge-based approach of sentiment analysis in major conflict ad hoc discussions of the social network Twitter. An experiment is made in which the quality of the proposed method is evaluated with different parameters on two real ad hoc discussions: Ferguson unrest (USA) and Biryuliovo bashings (Russia). The results of the experiment show a good quality of the sentiment analysis of the discussions. In particular, the average value of the accuracy of Russian and English is 0.65, and the f-measure is 0.7.