It would be helpful to consider various topic-independent features: syntax, semantics, and discourserelations between text fragments to accurately detect texts containing elements of hatred or enmity.Unfortunately, methods for identifying discourse relations in the texts of social networks are poorlydeveloped. The paper considers the task of classification of discourse relations between two parts ofthe text. The RST Discourse Treebank dataset (LDC2002T07) is used to assess the performance of themethods. Since the size of this dataset is too small for training large language models, the work uses amodel-pre fitting approach. Model pre-fitting is performed on a Reddit user comment dataset. Textsfrom this dataset are labeled automatically. Since automatic labeling is less accurate than manualmarking, we use the multiple-instance learning (MIL) method to train models. A distinctive feature ofmodern language models is the large number of parameters. Using several models at different levels ofsuch a text analyzer requires a lot of resources. Therefore, for the analyzer to work, it is necessary touse high-performance or distributed computing. The use of desktop grid systems can attract andcombine computing resources to solve this type of problem.