These docs are for version 1.x of Rasa Open Source. - [Docs for the new version 2.0 can be found here.](/content/docs/rasa/index.html) ## User Guide - [Installation](https://legacy-docs-v1.rasa.com/1.7.2/user-guide/installation/) - [Tutorial: Rasa Basics](https://legacy-docs-v1.rasa.com/1.7.2/user-guide/rasa-tutorial/) - [Tutorial: Building Assistants](https://legacy-docs-v1.rasa.com/1.7.2/user-guide/building-assistants/) - [Command Line Interface](https://legacy-docs-v1.rasa.com/1.7.2/user-guide/command-line-interface/) - [Architecture](https://legacy-docs-v1.rasa.com/1.7.2/user-guide/architecture/) - [Messaging and Voice Channels](https://legacy-docs-v1.rasa.com/1.7.2/user-guide/messaging-and-voice-channels/) - [Evaluating Models](https://legacy-docs-v1.rasa.com/1.7.2/user-guide/evaluating-models/) - [Validate Data](https://legacy-docs-v1.rasa.com/1.7.2/user-guide/validate-files/) - [Configuring the HTTP API](https://legacy-docs-v1.rasa.com/1.7.2/user-guide/configuring-http-api/) - [Deploying your Rasa Assistant](https://legacy-docs-v1.rasa.com/1.7.2/user-guide/how-to-deploy/) - [Cloud Storage](https://legacy-docs-v1.rasa.com/1.7.2/user-guide/cloud-storage/) ## NLU - [About](https://legacy-docs-v1.rasa.com/1.7.2/nlu/about/) - [Using NLU Only](https://legacy-docs-v1.rasa.com/1.7.2/nlu/using-nlu-only/) - [Training Data Format](https://legacy-docs-v1.rasa.com/1.7.2/nlu/training-data-format/) - [Choosing a Pipeline](https://legacy-docs-v1.rasa.com/1.7.2/nlu/choosing-a-pipeline/) - [Language Support](https://legacy-docs-v1.rasa.com/1.7.2/nlu/language-support/) - [Entity Extraction](https://legacy-docs-v1.rasa.com/1.7.2/nlu/entity-extraction/) - [Components](https://legacy-docs-v1.rasa.com/1.7.2/nlu/components/) ## Core - [About](https://legacy-docs-v1.rasa.com/1.7.2/core/about/) - [Stories](https://legacy-docs-v1.rasa.com/1.7.2/core/stories/) - [Domains](https://legacy-docs-v1.rasa.com/1.7.2/core/domains/) - [Responses](https://legacy-docs-v1.rasa.com/1.7.2/core/responses/) - [Actions](https://legacy-docs-v1.rasa.com/1.7.2/core/actions/) - [Policies](https://legacy-docs-v1.rasa.com/1.7.2/core/policies/) - [Slots](https://legacy-docs-v1.rasa.com/1.7.2/core/slots/) - [Forms](https://legacy-docs-v1.rasa.com/1.7.2/core/forms/) - [Retrieval Actions](https://legacy-docs-v1.rasa.com/1.7.2/core/retrieval-actions/) - [Interactive Learning](https://legacy-docs-v1.rasa.com/1.7.2/core/interactive-learning/) - [Fallback Actions](https://legacy-docs-v1.rasa.com/1.7.2/core/fallback-actions/) - [Knowledge Base Actions](https://legacy-docs-v1.rasa.com/1.7.2/core/knowledge-bases/) ## Conversation Design - [Dialogue Elements](https://legacy-docs-v1.rasa.com/1.7.2/dialogue-elements/dialogue-elements/) - [Small Talk](https://legacy-docs-v1.rasa.com/1.7.2/dialogue-elements/small-talk/) - [Completing Tasks](https://legacy-docs-v1.rasa.com/1.7.2/dialogue-elements/completing-tasks/) - [Guiding Users](https://legacy-docs-v1.rasa.com/1.7.2/dialogue-elements/guiding-users/) ## API Reference - [Action Server](https://legacy-docs-v1.rasa.com/1.7.2/api/action-server/) - [HTTP API](https://legacy-docs-v1.rasa.com/1.7.2/api/http-api/) - [Jupyter Notebooks](https://legacy-docs-v1.rasa.com/1.7.2/api/jupyter-notebooks/) - [Agent](https://legacy-docs-v1.rasa.com/1.7.2/api/agent/) - [Custom NLU Components](https://legacy-docs-v1.rasa.com/1.7.2/api/custom-nlu-components/) - [Rasa SDK](https://legacy-docs-v1.rasa.com/1.7.2/api/rasa-sdk/) - [Events](https://legacy-docs-v1.rasa.com/1.7.2/api/events/) - [Tracker](https://legacy-docs-v1.rasa.com/1.7.2/api/tracker/) - [Tracker Stores](https://legacy-docs-v1.rasa.com/1.7.2/api/tracker-stores/) - [Event Brokers](https://legacy-docs-v1.rasa.com/1.7.2/api/event-brokers/) - [Lock Stores](https://legacy-docs-v1.rasa.com/1.7.2/api/lock-stores/) - [Training Data Importers](https://legacy-docs-v1.rasa.com/1.7.2/api/training-data-importers/) - [Featurization of Conversations](https://legacy-docs-v1.rasa.com/1.7.2/api/core-featurization/#) - [Migration Guide](https://legacy-docs-v1.rasa.com/1.7.2/migration-guide/) - [Rasa OSS Change Log](https://legacy-docs-v1.rasa.com/1.7.2/changelog/) ## Migrate from (beta) - [Dialogflow](https://legacy-docs-v1.rasa.com/1.7.2/migrate-from/google-dialogflow-to-rasa/) - [Wit.ai](https://legacy-docs-v1.rasa.com/1.7.2/migrate-from/facebook-wit-ai-to-rasa/) - [LUIS](https://legacy-docs-v1.rasa.com/1.7.2/migrate-from/microsoft-luis-to-rasa/) - [IBM Watson](https://legacy-docs-v1.rasa.com/1.7.2/migrate-from/ibm-watson-to-rasa/) ## Reference - [Glossary](https://legacy-docs-v1.rasa.com/1.7.2/glossary/) ## Versions viewing: 1.7.2 **Warning:** This document is for an old version of Rasa. The latest version is 1.10.26. # Featurization of Conversations In order to apply machine learning algorithms to conversational AI, we need to build up vector representations of conversations. Each story corresponds to a tracker which consists of the states of the conversation just before each action was taken. ## State Featurizers Every event in a trackers history creates a new state (e.g. running a bot action, receiving a user message, setting slots). Featurizing a single state of the tracker has a couple steps: 1. **Tracker provides a bag of active features**: - 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` - features indicating what the last action was (e.g. `prev_action_listen`) 2. **Convert all the features into numeric vectors**: > We use the `X, y` notation that’s common for supervised learning, where `X` is an array of shape `(num_data_points, time_dimension, num_input_features)`, and `y` is an array of shape `(num_data_points, num_bot_features)` or `(num_data_points, time_dimension, num_bot_features)` containing the target class labels encoded as one-hot vectors. > The target labels correspond to actions taken by the bot. > To convert the features into vector format, there are different featurizers available: > - `BinarySingleStateFeaturizer` creates a binary one-hot encoding: > The vectors `X, y` indicate a presence of a certain intent, entity, previous action or slot e.g. `[0 0 1 0 0 1 ...]`. > - `LabelTokenizerSingleStateFeaturizer` creates a vector > based on the feature label: All active feature labels (e.g. `prev_action_listen`) are split into tokens and represented as a bag-of-words. For example, actions `utter_explain_details_hotel` and `utter_explain_details_restaurant` will have 3 features in common, and differ by a single feature indicating a domain. > Labels for user inputs (intents, entities) and bot actions are featurized separately. Each label in the two categories is tokenized on a special character `split_symbol` (e.g. `action_search_restaurant = {action, search, restaurant}`), creating two vocabularies. A bag-of-words representation is then created for each label using the appropriate vocabulary. The slots are featurized as binary vectors, indicating their presence or absence at each step of the dialogue. **Note** If the domain defines the possible `actions`, `[ActionGreet, ActionGoodbye]`, `4` additional default actions are added: `[ActionListen(), ActionRestart(), ActionDefaultFallback(), ActionDeactivateForm()]`. Therefore, label `0` indicates default action listen, label `1` default restart, label `2` a greeting and `3` indicates goodbye. ## Tracker Featurizers It’s often useful to include a bit more history than just the current state when predicting an action. The `TrackerFeaturizer` iterates over tracker states and calls a `SingleStateFeaturizer` for each state. There are two different tracker featurizers: ### 1. Full Dialogue `FullDialogueTrackerFeaturizer` creates numerical representation of stories to feed to a recurrent neural network where the whole dialogue is fed to a network and the gradient is backpropagated from all time steps. Therefore, `X` is an array of shape `(num_stories, max_dialogue_length, num_input_features)` and `y` is an array of shape `(num_stories, max_dialogue_length, num_bot_features)`. The smaller dialogues are padded with `-1` for all features, indicating no values for a policy. ### 2. Max History `MaxHistoryTrackerFeaturizer` creates an array of previous tracker states for each bot action or utterance, with the parameter `max_history` defining how many states go into each row in `X`. Deduplication is performed to filter out duplicated turns (bot actions or bot utterances) in terms of their previous states. Hence `X` has shape `(num_unique_turns, max_history, num_input_features)` and `y` is an array of shape `(num_unique_turns, num_bot_features)`. For some algorithms a flat feature vector is needed, so `X` should be reshaped to `(num_unique_turns, max_history * num_input_features)`. If numeric target class labels are needed instead of one-hot vectors, use `y.argmax(axis=-1)`. 👋 I can help you get started with Rasa and answer your technical questions.