Question and answer generation approaches focus on the quality and correctness of generated questions for online courses but miss to use a good question word, which is a deficiency reported by many previous studies. In this experimental setup, we explored whether the word2vec representation, which is semantic-based, can be used to predict question words. We compare two pipelines of the prediction process and observed that splitting the problem into several subproblems performs similar to feeding a neural network with all the data. Although our approach is promising to take the context-based representation into account we can see that the success rate is still low but better than guessing.