How is a training data set constructed from user questions for the IBM Watson Natural Language Classifier?