In this thesis we solved the problem of automatic resolving of good and bad characters in children's stories. This problem is comprised of recognition of characters in the text, and deciding for each found character whether he/she/it is good or bad. In the first part of the thesis we considered the theoretical background of this problem. We began by describing the field of natural language processing, which comprises many problems. Some of these problems we recognized as separate components of the character resolving problem. These are the problems of named entity recognition, coreference resolution and sentiment analysis. For each of these three problems, we described several machine learning models which are used to solve them. In the second part of the thesis we implemented a software solution for the character resolving problem. As a basis we used Stanford CoreNLP toolkit developed at Stanford University. Comprehensive documentation for the solution is given in the thesis, as well as detailed analysis of obtained results for several children's stories. Based on them we came to the conclusion about advantages and disadvantages of the implemented solution. Lastly, we gave several recommendations for improvement of the current solution, as well as for implementation of a new solution.