I have 10 classes and 10-15 questions in each class . Given a new question, I want to find the class to which the question is most similar?
This is called sequence classification. An approach to attack this problem is to deploy an LSTM-RNN. The key is the preprocessing of the dataset; how you encode a written question into appropriate input for the network.
The answer to this is called word embedding and detailed instructions of how to implement it can be found in this tutorial, hope it helps :)