Dialogue Policies

A dialogue policy is the strategy by which a bot selects its next action at each turn of a conversation. Bavard chatbots use three different types of dialogue policies:

  1. 1/ Conversation Flows

    Conversations flows, aka dialogue graphs, are simply flow charts. They control the bot's behavior exactly and are very useful for wide range of use cases. This type of dialogue policy is described in more detail here.

  2. AI based Dialogue

    With this dialogue policy, the chatbot is trained on a dataset of conversations to predict the next action based on the current conversation history.

    This type of dialogue policy is very powerful but requires more effort on the part of the bot's owners. They will need to create the set of training conversations that the bot will learn from, and the more training data the better. Simple behavior can be learned from small datasets of only a handful or a few dozen conversations. But to realise the full capabilities of modern NLP, large datasets of hundreds or even thousands of conversations may be necessary, especially for complicated uses cases involving multiple domains and many different agent action types.

  3. AI based Default Actions

    The default action policy, like the AI based dialogue, is based on machine learning. However it only make use of the most recent user utterance, that is, under this policy we have a single-turn chatbot that doesn't use any context from previous turns in the conversation. Although it provides less rich behavior, this type of chatbot is easy to create and very useful for doing FAQs or for transitioning between graph modes and chitchat modes.