Skip to content

May 13th, 2020

Using Custom spaCy Components in Rasa

  • portrait of Vincent Warmerdam

    Vincent Warmerdam

In this guide we're going to show you how you can get a custom spaCy model working inside of Rasa on your local machine. The document does expect that you're already familiar with spaCy and Rasa. If you're not, feel free to check out the spaCy online course or spaCy introductory youtube series. The getting started guide for Rasa can be found here.

Note that this guide was written for Rasa version 1.10.0 with spaCy version 2.2.4.

Installation

If you want to follow along you'll need to clone the repository over here and install all the dependencies.

git clone https://github.com/RasaHQ/spaCy-integration-demo
make install

This repository represents a simple assistant that only needs to understand four intents. These are described in the nlu.md file;

## intent:greet
- hey
- hello
...

## intent:goodbye
- bye
- goodbye
...

## intent:bot_challenge
- are you a bot?
- are you a human?
...

## intent:talk_code
- i want to talk about python- How do you do inline delegates in vb.net like python
- Code to ask yes/no question in javascript
- Executing JavaScript from Flex: Is this javascript function dangerous?
- What does this python error mean? 
...

Note that this file only contains intents, we do not have any entities defined here.

There's also a config.yml file. Let's make sure it has these contents;

language: en

pipeline:
- name: SpacyNLP
  model: "en_core_web_sm"
- name: SpacyTokenizer
- name: SpacyEntityExtractor
- name: SpacyFeaturizer
  pooling: mean
- name: CountVectorsFeaturizer
  analyzer: char_wb
  min_ngram: 1
  max_ngram: 2
- name: DIETClassifier
  epochs: 1

policies:
  - name: MemoizationPolicy
  - name: KerasPolicy
  - name: MappingPolicy

Let's note a few things here;

  1. The first step in the pipeline tells us that we're going to use the en_core_web_sm model in spaCy. This is equivalent to calling spacy.load("en_core_web_sm") which means that you need to make sure that it is downloaded beforehand via python -m spacy download en_core_web_sm.
  2. Because we're using the spaCy model we now also have to use the tokenizer from spaCy. We do this in the second pipeline step.
  3. In the third step we're telling spaCy to detect entities on our behalf.
  4. In the fourth step we're telling spaCy to also generate word embeddings. We take the mean of all these embeddings such that a single array is passed to the later steps.
  5. In the next steps we generate some features using the CountVectorsFeaturizer that will be passed to the DIETClassifier. Since we're interested in showing the effect of the SpacyEntityExtractor we're only training the algorithm for 1 epoch because we are zooming in on the spaCy part.

We can train this pipeline and talk to it to see what the effect is. Let's say Hi I am Vincent from Amsterdam to this assistant.

> rasa train
> rasa shell nlu
Next message: 
  Hi I am Vincent from Amsterdam. 

When you run this, you'll notice in the output that both Vincent and Amsterdam have been detected as entities.

...
"entities": [
    {
      "entity": "PERSON",
      "value": "Vincent",
      "start": 9,
      "confidence": null,
      "end": 16,
      "extractor": "SpacyEntityExtractor"
    },
    {
      "entity": "GPE",
      "value": "Amsterdam",
      "start": 22,
      "confidence": null,
      "end": 31,
      "extractor": "SpacyEntityExtractor"
    }
...

The standard en_core_web_sm spaCy model supports some basic entities right out of the box. These include people (PERSON) as well as countries, cities and states (GPE). Note that the spaCy model did not get trained by our rasa train command. As far as Rasa is concerned spaCy is treated as a pretrained model.

Customisation

Let's create our own spaCy model now and add that to the pipeline. We'll keep it simple by only having a NER model that uses a pattern matcher but the general pattern will apply to more advanced spaCy models as well. The pattern matcher in spaCy works by declaring a collection of patterns that can be used to detect entities. There's an example below.

matcher-rules/proglang.jsonl

{"label":"PROGLANG","pattern":[{"LOWER":"golang"}]}
{"label":"PROGLANG","pattern":[{"LOWER":"go", "POS": {"NOT_IN": "VERB"}}]}
{"label":"PROGLANG","pattern":[{"LOWER":"sql"}]}
{"label":"PROGLANG","pattern":[{"LOWER":"python"}]}
{"label":"PROGLANG","pattern":[{"LOWER":{"REGEX":"(python\\d+\.?\\d*.?\\d*)"}}]}
{"label":"PROGLANG","pattern":[{"LOWER":"python"}, {"TEXT":{"REGEX":"(\\d+\.?\\d*.?\\d*)"}}]}
{"label":"PROGLANG","pattern":[{"LOWER": {"IN": ["node", "nodejs", "js", "javascript"]}}]}

Most of the patterns that we're detecting here are based on regex but not all. The nice thing about spaCy matching rules is that we can also use part of speech in these patterns. That allows us to detect "go" as a programming language, but only if "go" is not used as a verb.

Towards a Model

To generate the spaCy model using this file you can run a script (mkmodel.py) that contains the following content;

import pathlib

import spacy
from spacy.lang.en import English
from spacy.pipeline import EntityRuler

if __name__ == "__main__":
    path = pathlib.Path('matcher-rules/proglang.jsonl')
    # note that we could have also used `English()` as a starting point
    # if our matching rules weren't using part of speech 
    nlp = spacy.load("en_core_web_sm")

    # create a new rule based NER detector loading in settings from disk
    ruler = EntityRuler(nlp).from_disk(path)
    print(f"Will now create model for {path}.")

    # add the detector to the model
    nlp.add_pipe(ruler, name="proglang")

    # define the name of the model as a package
    nlp.meta["name"] = "proglang"
    # save the model to disk
    nlp.to_disk(nlp.meta["name"])
    print(f"spaCy model saved over at {nlp.meta['name']}.")

This script will look in the matcher-rules folder and it will pick up the .jsonl file that contain the rules for the EntityRuler. Once loaded it will construct a spaCy model and save it to disk. After saving it to disk, it is a good habit to make a proper package out of it so that your virtualenv is aware. Let's first make sure we have a spaCy model on disk.

> python mkmodel.py
Will now create model for matcher-rules/proglang.jsonl.
spaCy model saved over at proglang.

We now have a saved spaCy model on disk. We could already load it with spaCy by calling spacy.load("proglang") and that means that we could also refer to it in our config.yml. For local use this is fine but for production use-case it would be nicer to properly package the spaCy model. Let's run the commands for that.

> python -m spacy package proglang . --force
✔ Loaded meta.json from file
proglang/meta.json
✔ Successfully created package 'en_proglang-2.2.5'
en_proglang-2.2.5

This command creates a python package folder structure.

en_proglang-2.2.5
├── MANIFEST.in
├── en_proglang
│ ├── __init__.py
│ └── en_proglang-2.2.5
│ ├── meta.json
│ ├── ner
│ │ ├── cfg
│ │ ├── model
│ │ └── moves
│ ├── parser
│ │ ├── cfg
│ │ ├── model
│ │ └── moves
│ ├── proglang
│ │ ├── cfg
│ │ └── patterns.jsonl
│ ├── tagger
│ │ ├── cfg
│ │ ├── model
│ │ └── tag_map
│ ├── tokenizer
│ └── vocab
│ ├── key2row
│ ├── lexemes.bin
│ ├── lookups.bin
│ ├── strings.json
│ └── vectors
├── meta.json
└── setup.py

We can tell python to create a tar file that we can pip install.

> cd en_proglang-2.2.5
> python setup.py sdist 
> cd .. 

The en_proglang-2.2.5 folder now has different contents.

en_proglang-2.2.5
├── MANIFEST.in
├── dist
│ └── en_proglang-2.2.5.tar.gz
├── en_proglang
│ ├── __init__.py
│ ├── en_proglang-2.2.5
│ │ ├── meta.json
│ │ ├── ner
│ │ │ ├── cfg
│ │ │ ├── model
│ │ │ └── moves
│ │ ├── parser
│ │ │ ├── cfg
│ │ │ ├── model
│ │ │ └── moves
│ │ ├── proglang
│ │ │ ├── cfg
│ │ │ └── patterns.jsonl
│ │ ├── tagger
│ │ │ ├── cfg
│ │ │ ├── model
│ │ │ └── tag_map
│ │ ├── tokenizer
│ │ └── vocab
│ │ ├── key2row
│ │ ├── lexemes.bin
│ │ ├── lookups.bin
│ │ ├── strings.json
│ │ └── vectors
│ └── meta.json
├── en_proglang.egg-info
│ ├── PKG-INFO
│ ├── SOURCES.txt
│ ├── dependency_links.txt
│ ├── not-zip-safe
│ ├── requires.txt
│ └── top_level.txt
├── meta.json
└── setup.py

But we can now safely install the model as a package.

> python -m pip install en_proglang-2.2.5/dist/en_proglang-2.2.5.tar.gz

By doing this we can now load the model in two ways from python.

> python 
>>> import spacy 
>>> spacy.load("en_proglang")
<spacy.lang.en.English object at 0x119d0b080>
>>> import en_proglang
>>> en_proglang.load()
<spacy.lang.en.English object at 0x119d593c8>

Configure

Now that this is packaged up we can refer to it in our config.yml. So here's one that refers to the en_proglang link we just made.

pipeline:
- name: SpacyNLP
  model: "en_proglang"
- name: SpacyTokenizer
- name: SpacyEntityExtractor
- name: SpacyFeaturizer
  pooling: mean
- name: CountVectorsFeaturizer
  analyzer: char_wb
  min_ngram: 1
  max_ngram: 4
- name: DIETClassifier
  epochs: 1

You'll notice that the config.yml file has a reference to en_proglang. This is equivalent to running spacy.load("en_proglang") and because it is a package we don't need to worry about filepaths. Now this model will be used for entity detection.

With these changes you can rerun the same experiment and now detect programming languages in the text.

> rasa train; rasa shell nlu
"i program using go"              # [go] is now a PROGLANG entity
"i want to talk about python 3.6" # [python 3.6] is now a PROGLANG entity
"i code with node"                # [node] is now a PROGLANG entity
"i live in Amsterdam"             # [Amsterdam] is now a GPE entity

This works but maybe we'd like to limit the entities here. We're only interested in the entities that refer to programming languages and currently the base model in spaCy is detecting it as an organisation. There's a few options here;

  1. we can change the spaCy model and turn off the native models, this would also make the pipeline faster
  2. we can change the spaCy model and have it use a better (but heavier) english model like en_core_web_lg as a starting point
  3. we can also just turn off the base entities from from config.yml

Let's do the latter option.

pipeline:
- name: SpacyNLP
  model: "en_proglang"
- name: SpacyTokenizer
- name: SpacyFeaturizer
  pooling: mean
- name: SpacyEntityExtractor
  dimensions: ["PROGLANG"]
- name: CountVectorsFeaturizer
  analyzer: char_wb
  min_ngram: 1
  max_ngram: 4
- name: DIETClassifier
  epochs: 1

We've added a dimensions property to SpacyEntityExtractor which will ensure that we only get entities that we ask for.

With this configuration, you should now be able to see new behavior.

> rasa train; rasa shell nlu
"i program using go"              # [go] is now a PROGLANG entity
"i want to talk about python 3.6" # [python 3.6] is now a PYTHON entity
"i code with node"                # [node] is now a JAVASCRIPT entity
"i live in Amsterdam"             # no entity detected

Use Cases

We've only scratched the surface of what is possible with spaCy but hopefully this guide was able to show you how you can connect a custom spaCy model to Rasa.

So you might wonder, when might this be useful? There's a few instances;

  • spaCy has an awesome suite of tools to detect entities and it may just be that your usecase fits their toolchain really well (like the pattern match for go, when it is not a verb it may be a programming language)
  • spaCy has support models in many languages, including some multi-language models, so if your assistant needs to speak Dutch, you could use a pretrained spaCy model for that while still using the other tools.
  • spaCy has pre-trained models that automatically have support for common entities such as people and places, meaning you don't need to train your own
  • spaCy has a large community of specialized pretrained models that you can download, say on legal texts or academic research papers

Not a Use Case

That said, you may not need it all the time.

The spaCy workflow can be great if you have a highly customized model and you'd like to get it into Rasa. But it may not be an ideal starting point though since spaCy is a tool for general NLP tasks while the tools that Rasa offers are in general more specialized for the digital assistant use case.

Play

Feel free to play around with this! You can change the empty starting model in mkmodel.py with a beefy pretrained. For the english language en_core_web_lg is a popular choice but there's even multi-lingual models to pick from.

Happy hacking!