Hi Pavel, I share your ideas for 2017, but I would like to explore also unsupervised learning for intent classification and the possibility to transfer knowledge across chatbots.
2016 was quite productive for me too. I have designed and implemented a framework for task-based chatbots, combining machine learning and rule-based approaches. In particular, in my implementation ML-based intent classification relies on word embedding and CNN built on Google Tensorflow.
Such approach allowed me to mitigate the “cold start” issue arising where labeled dataset is too small to train a deep learning model.
But it’s not enough. I’m now working on further reducing chatbot set-up time by introducing unsupervised and/or reinforcement learning (less labeled data), as well as maximizing the re-use of pre-trained models (new chatbots inherit knowledge from previous generations!).