Warning: This document is for an old version of rasa NLU. The latest version is 0.15.1.

Entity Extraction

There are a number of different entity extraction components, which can seem intimidating for new users. Here we’ll go through a few use cases and make recommendations of what to use.

Component Requires Model notes
ner_mitie MITIE structured SVM good for training custom entities
ner_crf pycrfsuite conditional random field good for training custom entities
ner_spacy spaCy averaged perceptron provides pre-trained entities
ner_duckling duckling context-free grammar provides pre-trained entities


To use these components, you will probably want to define a custom pipeline, see Processing Pipeline. You can add multiple ner components to your pipeline; the results from each will be combined in the final output.

Use Cases

Here we’ll outline some common use cases for entity extraction, and make recommendations on which components to use.

Places, Dates, People, Organisations

spaCy has excellent pre-trained named-entity recognisers in a number of models. You can test them out in this awesome interactive demo. We don’t recommend that you try to train your own NER using spaCy, unless you have a lot of data and know what you are doing. Note that some spaCy models are highly case-sensitive.

Dates, Amounts of Money, Durations, Distances, Ordinals

The duckling package does a great job of turning expressions like “next Thursday at 8pm” into actual datetime objects that you can use. It can also handle durations like “two hours”, amounts of money, distances, etc. Fortunately, there is also a python wrapper for duckling! You can use this component by installing the duckling package from PyPI and adding ner_duckling to your pipeline.

Custom, Domain-specific entities

In the introductory tutorial we build a restaurant bot, and create custom entities for location and cuisine. The best components for training these domain-specific entity recognisers are the ner_mitie and ner_crf components. It is recommended that you experiment with both of these to see what works best for your data set.