I want to use cosine similarity to identify the intent and pass it to RASA Core

Question

I want to use cosine similarity to identify the intent and pass it to RASA Core. In other words, I want to replace the NLU part with some other similarity calculation method. How to do it?

Answer 1

Currently, there is four classifiers implemented in Rasa-NLU:

• sklearn_intent_classifier
• mitie_intent_classifier
• keyword_intent_classifier
• embedding_intent_classifier

If you use embedding_intent_classifier.py by default it is used cosine similarity:

"similarity_type": 'cosine',  # string 'cosine' or 'inner'

How to customize your pipeline?

language: "en"

pipeline:
- name: "tokenizer_whitespace"
- name: "ner_crf"
- name: "ner_synonyms"
- name: "intent_featurizer_count_vectors"
- name: "intent_classifier_tensorflow_embedding"

See here for more details.

How to define my own Components?

Inherit from parent object Component and implement your own. If you need to define tfidf and cosine read here, and then compare your code with here.

from rasa_nlu.components import Component

class MyComponent(Component):
 def __init__(self, component_config=None):
     pass

 def train(self, training_data, cfg, **kwargs):
     pass

 def process(self, message, **kwargs):
     pass

 def persist(self, model_dir):
     pass

 @classmethod
 def load(cls, model_dir=None, model_metadata=None, cached_component=None,
          **kwargs):

Also do not forget to add it into pipeline:

pipeline:
- name: "MyComponent"