Version: Latest

Stories

Stories are a type of training data used to train your assistant's dialogue management model. Stories can be used to train models that are able to generalize to unseen conversation paths.

NLU-based assistants

This section refers to building NLU-based assistants. If you are working with Conversational AI with Language Models (CALM), this content may not apply to you.

Format

A story is a representation of a conversation between a user and an AI assistant, converted into a specific format where user inputs are expressed as intents (and entities when necessary), while the assistant's responses and actions are expressed as action names.

Here's an example of a dialogue in the Rasa story format:

stories:
- story: collect restaurant booking info # name of the story - just for debugging
steps:
- intent: greet # user message with no entities
- action: utter_ask_howcanhelp
- intent: inform # user message with entities
entities:
- location: "rome"
- price: "cheap"
- action: utter_on_it # action that the bot should execute
- action: utter_ask_cuisine
- intent: inform
entities:
- cuisine: "spanish"
- action: utter_ask_num_people

User Messages

While writing stories, you do not have to deal with the specific contents of the messages that the users send. Instead, you can take advantage of the output from the NLU pipeline, which lets you use just the combination of an intent and entities to refer to all the possible messages the users can send to mean the same thing.

It is important to include the entities here as well because the policies learn to predict the next action based on a combination of both the intent and entities (you can, however, change this behavior using the use_entities attribute).

Actions

All actions executed by the bot, including responses are listed in stories under the action key.

You can use a response from your domain as an action by listing it as one in a story. Similarly, you can indicate that a story should call a custom action by including the name of the custom action from the actions list in your domain.

Events

During training, Rasa does not call the action server. This means that your assistant's dialogue management model doesn't know which events a custom action will return.

Because of this, events such as setting a slot or activating/deactivating a form have to be explicitly written out as part of the stories. For more info, see the documentation on Events.

Slot Events

Slot events are written under slot_was_set in a story. If this slot is set inside a custom action, add the slot_was_set event immediately following the custom action call. If your custom action resets a slot value to None, the corresponding event for that would look like this:

stories:
- story: set slot to none
steps:
# ... other story steps
- action: my_custom_action
- slot_was_set:
- my_slot: null

Form Events

There are three kinds of events that need to be kept in mind while dealing with forms in stories.

  • A form action event (e.g. - action: restaurant_form) is used in the beginning when first starting a form, and also while resuming the form action when the form is already active.

  • A form activation event (e.g. - active_loop: restaurant_form) is used right after the first form action event.

  • A form deactivation event (e.g. - active_loop: null), which is used to deactivate the form.

writing form stories

In order to get around the pitfall of forgetting to add events, the recommended way to write these stories is to use interactive learning.

Checkpoints and OR statements

Checkpoints and OR statements should be used with caution, if at all. There is usually a better way to achieve what you want by using Rules or the ResponseSelector.

Checkpoints

You can use checkpoints to modularize and simplify your training data. Checkpoints can be useful, but do not overuse them. Using lots of checkpoints can quickly make your example stories hard to understand, and will slow down training.

Here is an example of stories that contain checkpoints:

stories:
- story: beginning of flow
steps:
- intent: greet
- action: action_ask_user_question
- checkpoint: check_asked_question
- story: handle user affirm
steps:
- checkpoint: check_asked_question
- intent: affirm
- action: action_handle_affirmation
- checkpoint: check_flow_finished
- story: handle user deny
steps:
- checkpoint: check_asked_question
- intent: deny
- action: action_handle_denial
- checkpoint: check_flow_finished
- story: finish flow
steps:
- checkpoint: check_flow_finished
- intent: goodbye
- action: utter_goodbye
note

Unlike regular stories, checkpoints are not restricted to starting with user input. As long as the checkpoint is inserted at the right points in the main stories, the first event can be a custom action or a response as well.

Or Statements

Another way to write shorter stories, or to handle multiple intents or slot events the same way, is to use an or statement. For example, if you ask the user to confirm something, and you want to treat the affirm and thankyou intents in the same way. The story below will be converted into two stories at training time:

stories:
- story:
steps:
# ... previous steps
- action: utter_ask_confirm
- or:
- intent: affirm
- intent: thankyou
- action: action_handle_affirmation

You can also use or statements with slot events. The following means the story requires that the current value for the name slot is set and is either joe or bob:

stories:
- story:
steps:
- intent: greet
- action: utter_greet
- intent: tell_name
- or:
- slot_was_set:
- name: joe
- slot_was_set:
- name: bob
# ... next actions

or statements can be useful, but if you are using a lot of them, it is probably better to restructure your domain and/or intents. Overusing OR statements will slow down training.

Test Conversation Format

The test conversation format is a format that combines both NLU data and stories into a single file for evaluation. Read more about this format in Testing Your Assistant.

testing only

This format is only used for testing and cannot be used for training.

End-to-end Training

New in 2.2

End-to-end training is an experimental feature. We introduce experimental features to get feedback from our community, so we encourage you to try it out! However, the functionality might be changed or removed in the future. If you have feedback (positive or negative) please share it with us on the Rasa Forum.

With end-to-end training, you do not have to deal with the specific intents of the messages that are extracted by the NLU pipeline or with separate utter_ responses in the domain file. Instead, you can include the text of the user messages and/or bot responses directly in your stories. See the training data format for detailed description of how to write end-to-end stories.

You can mix training data in the end-to-end format with labeled training data which has intents and actions specified: Stories can have some steps defined by intents/actions and other steps defined directly by user or bot utterances.

We call it end-to-end training because policies can consume and predict actual text. For end-to-end user inputs, intents classified by the NLU pipeline and extracted entities are ignored.

Only Rule Policy and TED Policy allow end-to-end training.

  • RulePolicy uses simple string matching during prediction. Namely, rules based on user text will only match if the user text strings inside your rules and input during prediction are identical.

  • TEDPolicy passes user text through an additional Neural Network to create hidden representations of the text. In order to obtain robust performance you need to provide enough training stories to capture a variety of user texts for any end-to-end dialogue turn.

Rasa policies are trained for next utterance selection. The only difference to creating utter_ response is how TEDPolicy featurizes bot utterances. In case of an utter_ action, TEDPolicy sees only the name of the action, while if you provide actual utterance using bot key, TEDPolicy will featurize it as textual input depending on the NLU configuration. This can help in case of similar utterances in slightly different situations. However, this can also make things harder to learn because the fact that different utterances have similar texts make it easier for TEDPolicy to confuse these utterances.

End-to-end training requires significantly more parameters in TEDPolicy. Therefore, training an end-to-end model might require significant computational resources depending on how many end-to-end turns you have in your stories.