Training Conversations

To create a conversational assistant that is able to complete specific tasks using the full context of a conversation, you must train your assistant on a dataset of example conversations. These training conversations are a sequence of dialogue turns where each turn is annotated with additional pieces of infomation like the intent, tags, action type, and available slots. The specific annotations for each turn depend on whether it is a user turn or an agent turn.

Below is a very basic example of a training conversation. It teaches the bot to ask the user's name and then greet the user name.

User Turns

INTENT:GREET
UTTERANCE: Hi
INTENT: INFORM_NAME
UTTERANCE: My name is David.

Assistant Turns

ACTION:GREET
UTTERANCE: Hi, I'm a Bavard chatbot.
ACTION: ASK_NAME
UTTERANCE: What is your name?
ACTION: GREET_BY_NAME
UTTERANCE: Nice to meet you, {{ PERSON }}.

In the last turn, you'll notice that the assistant utterance uses a template where text {{ PERSON }} will be replaced with the value of the PERSON slot.