For many students, course recommendation is a troublesome problem. To choose satisfying courses when they only have a little of information about the courses, they have to turn to common course schedule systems for help, but the result is disappointing. To solve the problem, the method of scoring similarity of courses and clustering courses based on course descriptions is presented. Different from the method, this paper applies semantic similarity analysis into course selection, realizes a course recommendation system. Each course description is first modeled as a document, and a cleantokenize-TFIDF-Word2Vec-Doc2Vec pipeline is built to create vectors for each course from which cosine similarities will be calculated. The result is even better than the above method through evaluation.