Why we chose to work in rasa core with deep learning not simply machine learning to predict bot’s responses ? Is it because the memorization feature, so we need LSTM ? Is there any existing article defending the choice of deep learning and the used models(policies) ?