A library of useful custom Rasa components


Keywords
rasa, nlu, components, bert, bert-model, document-embeddings, entity-extraction, natural-language-processing, natural-language-understanding, nlp, nlp-machine-learning, rasa-nlu, universal-sentence-encoder
License
Apache-2.0
Install
pip install innatis==0.5.6

Documentation

Innatis

This is a library of custom Rasa NLU components that we (CarLabs) are building.

What does the name mean?

Viribus Innatis means "innate abilities" in Latin. It's a joke...

"Rasa" comes from "tabula rasa" - blank slate in Latin. We (just Sam) thought it would be funny for this project to have the opposite name, since this is meant to be a suite of tools to fill in functionality for Rasa... that is, make it a not-blank-slate. So "Innate Abilities" -> Viribus Innatis.

Usage

$ pip install innatis

Then add to your pipeline in your rasa_config.yml. Example pipeline can be found in sample_rasa_innatis_config.yml.

Components

Classifiers

  • intent_classifier_bert - Pulls the bert model from TF HUB and pretrains on given data.

Example config

language: en
pipeline:
  - name: "tokenizer_whitespace"
  - name: "ner_crf"
  - name: "ner_synonyms"
  - name: "innatis.classifiers.BertIntentClassifier"
    pretrained_model_dir: '/path/to/uncased_L-24_H-1024_A-16'
    epochs: 10
    batch_size: 64

Extractors

  • composite_entity_extractor - Given entities extracted by another extractor (ner_crf seems to be the best for now), splits them into composite entities, similar to DialogFlow.
  • EntitySynonymMapper (replaces ner_synonyms) - this is the ner_synonyms adapted for composite_entity_extractor. You most likely need it if you use composite_entity_extractor. It replaces the synonyms with the original entities inside composite entities. It can also do fuzzy matching when matching synonyms (enabled by default). See example config: config_composite_entities.yml

Featurizers

  • universal_sentence_encoder_featurizer - Pulls the smaller USE model from TF HUB and embeds inputs as document vectors, and that vector gets sent downstream to be used as a feature.

Development

git clone git@github.com:Revmaker/innatis.git
cd innatis
pipenv install
# pipenv install --skip-lock if locking takes too long

# do some stuff
# write some tests

pipenv run python test.py

Dependencies suck, as is always the case in Python. I started out using pipenv because I thought that was the package manager for Python. But I guess you need to put dependencies in the install_requires section of setup.py. So, the (currently manual, but automatable) process for moving dependencies from the Pipfile to setup.py is:

$ cd innatis
# you are now in parent/innatis, not innatis/innatis
$ python ir_from_pipfile.py | pbcopy
['scikit-learn', 'scipy', 'sklearn-crfsuite', 'tensorflow', 'word2number', 'rasa_nlu==0.13.8', 'tensorflow-hub', 'spacy']

# that array is copied; paste it into setup.py

Also, please manually bump the version in a semver-ish way.