Paragraph classification is possible only when natural language processing technologies such as morpheme analysis, classification analysis, and semantic analysis are combined. Therefore, implementing paragraph classification in a rule-based system is difficult, and performance is difficult. This problem is to be solved through a machine learning system. The machine learning system was designed using word similarity as a key technology. This system is divided into a system that processes text data into data suitable for learning and a system that performs machine learning. Three variables were set and tested to find the case of the best performance of this system. The variables set are the number of sentences to be compared at the same time, the size of the padding for parallel operations, and the optimizer to be used. Through experiments, the environment with the highest performance was found and compared with existing rule-based systems based on this. For accurate performance comparison, the same input data were used and F1-Score was selected as an evaluation indicator. As a result of the comparison, 15% performance improvement was confirmed.