This is unreleased documentation for Rasa Documentation Main/Unreleased version.
For the latest released documentation, see the latest version (3.x).

Version: Main/Unreleased

Generating NLU Data

NLU (Natural Language Understanding) is the part of Rasa that performs intent classification, entity extraction, and response retrieval.

NLU will take in a sentence such as "I am looking for a French restaurant in the center of town" and return structured data like:

"intent": "search_restaurant",
"entities": {
"cuisine": "French",
"location": "center"

Building NLU models is hard, and building ones that are production-ready is even harder. Here are some tips for designing your NLU training data and pipeline to get the most out of your bot.

Conversation-Driven Development for NLU

Conversation-Driven Development (CDD) means letting real user conversations guide your development. For building a great NLU model, this means two key things:

Gather Real Data

When it comes to building out NLU training data, developers are sometimes tempted to use text generation tools or templates to quickly increase the number of training examples. This is a bad idea for two reasons:

  • First, your synthetic data won't look like the messages that users actually send to your assistant, so your model will underperform.
  • Second, by training and testing on synthetic data, you trick yourself into thinking that your model is actually performing well, and you won't notice major issues.

Remember that if you use a script to generate training data, the only thing your model can learn is how to reverse-engineer the script.

To avoid these problems, it is always a good idea to collect as much real user data as possible to use as training data. Real user messages can be messy, contain typos, and be far from 'ideal' examples of your intents. But keep in mind that those are the messages you're asking your model to make predictions about! Your assistant will always make mistakes initially, but the process of training & evaluating on user data will set your model up to generalize much more effectively in real-world scenarios.

Share with Test Users Early

In order to gather real data, you’re going to need real user messages. A bot developer can only come up with a limited range of examples, and users will always surprise you with what they say. This means you should share your bot with test users outside the development team as early as possible. See the full CDD guidelines for more details.

Avoiding Intent Confusion

Intents are classified using character and word-level features extracted from your training examples, depending on what featurizers you've added to your NLU pipeline. When different intents contain the same words ordered in a similar fashion, this can create confusion for the intent classifier.

Splitting on Entities vs Intents

Intent confusion often occurs when you want your assistant's response to be conditioned on information provided by the user. For example, "How do I migrate to Rasa from IBM Watson?" versus "I want to migrate from Dialogflow."

Since each of these messages will lead to a different response, your initial approach might be to create separate intents for each migration type, e.g. watson_migration and dialogflow_migration. However, these intents are trying to achieve the same goal (migrating to Rasa) and will likely be phrased similarly, which may cause the model to confuse these intents.

To avoid intent confusion, group these training examples into single migration intent and make the response depend on the value of a categorical product slot that comes from an entity. This also makes it easy to handle the case when no entity is provided, e.g. "How do I migrate to Rasa?" For example:

- story: migrate from IBM Watson
- intent: migration
- product
- slot_was_set:
- product: Watson
- action: utter_watson_migration
- story: migrate from Dialogflow
- intent: migration
- product
- slot_was_set:
- product: Dialogflow
- action: utter_dialogflow_migration
- story: migrate from unspecified
- intent: migration
- action: utter_ask_migration_product

Improving Entity Recognition

With Rasa, you can define custom entities and annotate them in your training data to teach your model to recognize them. Rasa also provides components to extract pre-trained entities, as well as other forms of training data to help your model recognize and process entities.

Pre-trained Entity Extractors

Common entities such as names, addresses, and cities require a large amount of training data for an NLU model to generalize effectively.

Rasa provides two great options for pre-trained extraction: SpacyEntityExtractor and DucklingEntityExtractor. Because these extractors have been pre-trained on a large corpus of data, you can use them to extract the entities they support without annotating them in your training data.


Regexes are useful for performing entity extraction on structured patterns such as 5-digit U.S. zip codes. Regex patterns can be used to generate features for the NLU model to learn, or as a method of direct entity matching. See Regular Expression Features for more information.

Lookup Tables

Lookup tables are processed as a regex pattern that checks if any of the lookup table entries exist in the training example. Similar to regexes, lookup tables can be used to provide features to the model to improve entity recognition, or used to perform match-based entity recognition. Examples of useful applications of lookup tables are flavors of ice cream, brands of bottled water, and even sock length styles (see Lookup Tables).


Adding synonyms to your training data is useful for mapping certain entity values to a single normalized entity. Synonyms, however, are not meant for improving your model's entity recognition and have no effect on NLU performance.

A good use case for synonyms is when normalizing entities belonging to distinct groups. For example, in an assistant that asks users what insurance policies they're interested in, they might respond with "my truck," "a car," or "I drive a batmobile." It would be a good idea to map truck, car, and batmobile to the normalized value auto so that the processing logic will only need to account for a narrow set of possibilities (see synonyms).

Synonyms can also be used to standardize the extracted entities. A synonym for iPhone can map iphone or IPHONE to the synonym without adding these options in the synonym examples.

Handling Edge Cases


Coming across misspellings is inevitable, so your bot needs an effective way to handle this. Keep in mind that the goal is not to correct misspellings, but to correctly identify intents and entities. For this reason, while a spellchecker may seem like an obvious solution, adjusting your featurizers and training data is often sufficient to account for misspellings.

Adding a character-level featurizer provides an effective defense against spelling errors by accounting for parts of words, instead of only whole words. You can add character level featurization to your pipeline by using the char_wb analyzer for the CountVectorsFeaturizer, for example:

# <other components>
- name: CountVectorsFeaturizer
analyze: char_wb
min_ngram: 1
max_ngram: 4
# <other components>

In addition to character-level featurization, you can add common misspellings to your training data.

Defining an Out-of-scope Intent

It is always a good idea to define an out_of_scope intent in your bot to capture any user messages outside of your bot's domain. When an out_of_scope intent is identified, you can respond with messages such as "I'm not sure how to handle that, here are some things you can ask me..." to gracefully guide the user towards a supported skill.

Shipping Updates

Treat your data like code. In the same way that you would never ship code updates without reviews, updates to your training data should be carefully reviewed because of the significant influence it can have on your model's performance.

Use a version control system such as Github or Bitbucket to track changes to your data and rollback updates when necessary.

Be sure to build tests for your NLU models to evaluate performance as training data and hyper-parameters change. Automate these tests in a CI pipeline such as Jenkins or Git Workflow to streamline your development process and ensure that only high-quality updates are shipped.