Custom NLU Components
Custom NLU Components
You can create a custom component to perform a specific task which NLU doesn’t currently offer (for example, sentiment analysis). Below is the specification of the rasa.nlu.components.Component class with the methods you’ll need to implement.
Note
There is a detailed tutorial on building custom components here.
You can add a custom component to your pipeline by adding the module path.
So if you have a module called sentiment containing a SentimentAnalyzer class:
pipeline: - name: "sentiment.SentimentAnalyzer"
Also be sure to read the section on the Component Lifecycle.
To get started, you can use this skeleton that contains the most important methods that you should implement:
<br> 1<br> 2<br> 3<br> 4<br> 5<br> 6<br> 7<br> 8<br> 9<br>10<br>11<br>12<br>13<br>14<br>15<br>16<br>17<br>18<br>19<br>20<br>21<br>22<br>23<br>24<br>25<br>26<br>27<br>28<br>29<br>30<br>31<br>32<br>33<br>34<br>35<br>36<br>37<br>38<br>39<br>40<br>41<br>42<br>43<br>44<br>45<br>46<br>47<br>48<br>49<br>50<br>51<br>52<br>53<br>54<br>55<br>56<br>57<br>58<br>59<br>60<br>61<br>62<br>63<br>64<br>65<br>66<br>67<br>68<br>69<br>70<br>71<br>72<br>73<br>74<br>75<br>76<br>77<br>78<br>79<br>80<br>81<br>82<br>83<br>84<br>85<br>86<br>87<br>88<br> |
<br>import typing<br>from typing import Any, Optional, Text, Dict, List, Type<br>from rasa.nlu.components import Component<br>from rasa.nlu.config import RasaNLUModelConfig<br>from rasa.nlu.training_data import Message, TrainingData<br>if typing.TYPE_CHECKING:<br> from rasa.nlu.model import Metadata<br>class MyComponent(Component):<br> """A new component"""<br> # Which components are required by this component.<br> # Listed components should appear before the component itself in the pipeline.<br> @classmethod<br> def required_components(cls) -> List[Type[Component]]:<br> """Specify which components need to be present in the pipeline."""<br> return []<br> # Defines the default configuration parameters of a component<br> # these values can be overwritten in the pipeline configuration<br> # of the model. The component should choose sensible defaults<br> # and should be able to create reasonable results with the defaults.<br> defaults = {}<br> # Defines what language(s) this component can handle.<br> # This attribute is designed for instance method: `can_handle_language`.<br> # Default value is None which means it can handle all languages.<br> # This is an important feature for backwards compatibility of components.<br> language_list = None<br> def __init__(self, component_config: Optional[Dict[Text, Any]] = None) -> None:<br> super().__init__(component_config)<br> def train(<br> self,<br> training_data: TrainingData,<br> config: Optional[RasaNLUModelConfig] = None,<br> **kwargs: Any,<br> ) -> None:<br> """Train this component.<br> This is the components chance to train itself provided<br> with the training data. The component can rely on<br> any context attribute to be present, that gets created<br> by a call to :meth:`components.Component.pipeline_init`<br> of ANY component and<br> on any context attributes created by a call to<br> :meth:`components.Component.train`<br> of components previous to this one."""<br> pass<br> def process(self, message: Message, **kwargs: Any) -> None:<br> """Process an incoming message.<br> This is the components chance to process an incoming<br> message. The component can rely on<br> any context attribute to be present, that gets created<br> by a call to :meth:`components.Component.pipeline_init`<br> of ANY component and<br> on any context attributes created by a call to<br> :meth:`components.Component.process`<br> of components previous to this one."""<br> pass<br> def persist(self, file_name: Text, model_dir: Text) -> Optional[Dict[Text, Any]]:<br> """Persist this component to disk for future loading."""<br> pass<br> @classmethod<br> def load(<br> cls,<br> meta: Dict[Text, Any],<br> model_dir: Optional[Text] = None,<br> model_metadata: Optional["Metadata"] = None,<br> cached_component: Optional["Component"] = None,<br> **kwargs: Any,<br> ) -> "Component":<br> """Load this component from file."""<br> if cached_component:<br> return cached_component<br> else:<br> return cls(meta)<br> |
Note
If you create a custom tokenizer you should implement the methods of rasa.nlu.tokenizers.tokenizer.Tokenizer.
The train and process methods are already implemented and you simply need to overwrite the tokenize method. train and process will automatically add a special token __CLS__ to the end of list of tokens, which is needed further down the pipeline.
Note
If you create a custom featurizer you should return a sequence of features.
E.g. your featurizer should return a matrix of size (number-of-tokens x feature-dimension). The feature vector of the __CLS__ token should contain features for the complete message.
Component
classrasa.nlu.components.Component( component_config=None)
A component is a message processing unit in a pipeline.
Components are collected sequentially in a pipeline. Each component is called one after another. This holds for initialization, training, persisting and loading the components.
If a component comes first in a pipeline, its methods will be called first.
E.g. to process an incoming message, the process method of each component will be called. During the processing (as well as the training, persisting and initialization) components can pass information to other components. The information is passed to other components by providing attributes to the so called pipeline context. The pipeline context contains all the information of the previous components a component can use to do its own processing. For example, a featurizer component can provide features that are used by another component down the pipeline to do intent classification.
classmethodrequired_components()
Specify which components need to be present in the pipeline.
Returns
The list of class names of required components.
Return type
List[Type[ Component]]
classmethodrequired_packages()
Specify which python packages need to be installed.
E.g. ['spacy']. More specifically, these should be importable python package names e.g. sklearn and not package names in the dependencies sense e.g. scikit-learn
This list of requirements allows us to fail early during training if a required package is not installed.
Returns
The list of required package names.
Return type
List[str]
classmethodcreate( component_config, config)
Creates this component (e.g. before a training is started).
Method can access all configuration parameters.
Parameters
- component_config – The components configuration parameters.
- config – The model configuration parameters.
Returns
The created component.
Return type
provide_context()
Initialize this component for a new pipeline.
This function will be called before the training is started and before the first message is processed using the interpreter. The component gets the opportunity to add information to the context that is passed through the pipeline during training and message parsing. Most components do not need to implement this method. It’s mostly used to initialize framework environments like MITIE and spacy (e.g. loading word vectors for the pipeline).
Returns
The updated component configuration.
Return type
Optional[Dict[str, Any]]
train( training_data, config=None, **kwargs)
Train this component.
This is the components chance to train itself provided with the training data. The component can rely on any context attribute to be present, that gets created by a call to rasa.nlu.components.Component.create() of ANY component and on any context attributes created by a call to rasa.nlu.components.Component.train() of components previous to this one.
Parameters
- training_data – The
rasa.nlu.training_data.training_data.TrainingData. - config – The model configuration parameters.
Return type
None
process( message, **kwargs)
Process an incoming message.
This is the components chance to process an incoming message. The component can rely on any context attribute to be present, that gets created by a call to rasa.nlu.components.Component.create() of ANY component and on any context attributes created by a call to rasa.nlu.components.Component.process() of components previous to this one.
Parameters
message – The rasa.nlu.training_data.message.Message to process.
Return type
None
persist( file_name, model_dir)
Persist this component to disk for future loading.
Parameters
- file_name – The file name of the model.
- model_dir – The directory to store the model to.
Returns
An optional dictionary with any information about the stored model.
Return type
Optional[Dict[str, Any]]
prepare_partial_processing( pipeline, context)
Sets the pipeline and context used for partial processing.
The pipeline should be a list of components that are previous to this one in the pipeline and have already finished their training (and can therefore be safely used to process messages).
Parameters
- pipeline – The list of components.
- context – The context of processing.
Return type
None
partially_process( message)
Allows the component to process messages during training (e.g. external training data).
The passed message will be processed by all components previous to this one in the pipeline.
Parameters
message – The rasa.nlu.training_data.message.Message to process.
Returns
The processed rasa.nlu.training_data.message.Message.
Return type
Message
classmethodcan_handle_language( language)
Check if component supports a specific language.
This method can be overwritten when needed. (e.g. dynamically determine which language is supported.)
Parameters
language – The language to check.
Returns
True if component can handle specific language, False otherwise.
Return type
bool