In order to apply machine learning algorithms to conversational AI, we need to build up vector representations of conversations.

We use the X, y notation that’s common for supervised learning, where X is a matrix of shape (num_data_points, data_dimension), and y is a 1D array of length num_data_points containing the target class labels.

The target labels correspond to actions taken by the bot. If the domain defines the possible actions [ActionGreet, ActionListen] then a label 0 indicates a greeting and 1 indicates a listen.

The rows in X correspond to the state of the conversation just before the action was taken.

Featurising a single state works like this: the tracker provides a bag of active_features comprising:

  • what the last action was (e.g. prev_action_listen)
  • features indicating intents and entities, if this is the first state in a turn, e.g. it’s the first action we will take after parsing the user’s message. (e.g. [intent_restaurant_search, entity_cuisine] )
  • features indicating which slots are currently defined, e.g. slot_location if the user previously mentioned the area they’re searching for restaurants.
  • features indicating the results of any API calls stored in slots, e.g. slot_matches
All of these features are represented in a binary vector which just indicates if they’re present.
e.g. [0 0 1 1 0 1 ...]

To recover the bag of features from a vector vec, you can call domain.reverse_binary_encoded_features(vec). This is very useful for debugging.


It’s often useful to include a bit more history than just the current state in memory. The parameter max_history defines how many states go into defining each row in X.

Hence the statement above that X is 2D is actually false, it has shape (num_states, max_history, num_features). For most algorithms you want a flat feature vector, so you will have to reshape this to (num_states, max_history * num_features).